Fe

dera

l Res

erve

Ban

k of

Chi

cago

Social interventions, health and wellbeing: The long-term and intergenerational effects of a school construction program

Bhashkar Mazumder, Maria Fernanda Rosales, and Margaret Triyana

REVISED

July 21, 2020

WP 2019-09

https://doi.org/10.21033/wp-2019-09

*Working papers are not edited, and all opinions and errors are the responsibility of the author(s). The views expressed do not necessarily reflect the views of the Federal Reserve Bank of Chicago or the Federal Reserve System.

Social interventions, health and wellbeing: The long-term and

intergenerational effects of a school construction program ∗

Bhashkar Mazumder† Maria Rosales-Rueda‡ Margaret Triyana§

This Draft: July 21, 2020

Abstract

We analyze the long-run and intergenerational effects of a large-scale school building

project (INPRES) that took place in Indonesia between 1974 and 1979. Specifically,

we link the geographic rollout of INPRES to longitudinal data from the Indonesian

Family Life Survey covering two generations. We find that individuals exposed to the

program have better health later in life along multiple measures. We also find that the

children of those exposed experience improved health and educational outcomes and

that these effects are generally stronger for maternal exposure than paternal exposure.

We find some evidence that household resources, neighborhood quality, and assortative

mating may explain a portion of our results. Our findings highlight the importance of

considering the long-run and multigenerational benefits when evaluating the costs and

benefits of social interventions in a middle-income country.

Keywords: Intergenerational transmission of human capital, education, adult well-

being

JEL Codes: J13, O15, I38

∗We thank Esther Duflo for sharing the administrative school construction data. We are grateful toPrashant Bharadwaj, Janet Currie, Taryn Dinkleman, Willa Friedman, Mike Grossman, Marco Gonzalez-Navarro, Joe Kaboski, Julien Labonne, Jean Lee, Elaine Liu, Grant Miller, Javaeria Qureshi, and seminarparticipants at the University of Notre Dame, Columbia and Princeton University Conference for Women inEconomics, American Economic Association Annual Meetings, Population Association of the Americas An-nual Meetings, NBER Development Meetings (Pre-conference), Millennium Challenge Corporation (MCC),LACEA-RIDGE conference, the NBER Health Economics Summer Institute, and the NBER Children SpringMeeting. We thank Leah Plachinski for excellent research assistance. We thank Dr. Amich Alhumami andDr. Abdul Malik for their helpful insight on the school construction program.†bhash.mazumder@gmail.com, Federal Reserve Bank of Chicago and University of Bergen‡maria.rosales@rutgers.edu, Rutgers University§mtriyana@alumni.uchicago.edu, Wake Forest University

1

1 Introduction

One of the fundamental ways that nations have tried to advance economic development

has been through investments in human capital (Becker, 1964). Many low and middle

income countries (LMICs) have embarked on educational reforms such as school construction

projects over the last fifty years in an effort to improve living standards. Indeed, one of the

prime UN Millennium Development Goals, listed second only to eradicating extreme poverty

and hunger, is to achieve universal primary education. While significant progress has been

made towards this goal and LMICs spend approximately one trillion dollars annually on

primary schooling, inadequate access to schools and poor school quality remain a pervasive

problem in much of the developing world (UNESCO, 2015). For example, primary schooling

rates in Senegal and Ethiopia were still below 60 percent in 2015.

For countries in the developing world that still need to embark on large-scale school

building initiatives, the potential societal gains may be significantly larger than current

research on educational policies may suggest. There is now increasing recognition that in

addition to improving purely economic outcomes, educational policies have the potential for

producing important spillovers on other aspects of well-being such as health (Galama et al.,

2018; Heckman et al., 2018; Oreopoulos and Salvanes, 2011). Furthermore, the human capital

gains in one generation may also be transmitted to the next generation, producing future

benefits that existing studies typically ignore. However, causal evidence of such spillovers and

intergenerational effects is limited, particularly in developing countries (Wantchekon et al.,

2014). A failure to properly consider these additional potential benefits could dramatically

understate the value of social interventions designed to improve human capital.

We focus on these non-pecuniary and intergenerational spillovers by studying a massive

primary school construction program in Indonesia known as the INPRES program. In partic-

ular, we examine the effects of this program on the long-term health of individuals who were

exposed to these schools (first generation) and the human capital outcomes of their offspring

(second generation). Under the INPRES program, the Indonesian government constructed

2

61,000 elementary schools between 1974 and 1979, doubling the existing stock of schools,

thus making schools available to children where none or few had existed before. Our study

builds upon earlier studies that have analyzed the effects of this program. Duflo (2001),

Duflo (2004), Breierova and Duflo (2004) and Zha (2019) show that INPRES affected the

education, marriage market, fertility, and labor-market outcomes of individuals exposed to

these schools, while Martinez-Bravo (2017) documents the impacts on local governance and

public good provision in Indonesia’s main island, Java. As may be expected, the main goal

of INPRES was to increase primary school attendance and in a companion paper (Mazumder

et al., 2019), we show that there were in fact large effects of the program on primary school

completion.

We build upon these studies and show that there were important spillovers on the long-

term health of those exposed to INPRES, as well as meaningful intergenerational effects.

Specifically, we use longitudinal data from the Indonesian Family Life Survey (IFLS), which

includes 5 waves between 1993 and 2014, and the Indonesian Family Life Survey-East (IFLS-

E), which includes one wave in 2012. These data allow us to track the offspring of INPRES

exposed individuals and examine important effects on the human capital outcomes of the

next generation. By using data explicitly designed to track an intergenerational sample,

our study is complementary to a contemporaneous paper by Akresh et al. (2018) who also

examine the long-run effects and intergenerational effects of INPRES. Akresh et al. (2018)

consider a different set of outcomes for adult children who are coresident with their parents

in cross-sectional survey data from 2016 (2016 Socio-economic survey, Susenas).1 Like the

earlier studies, we exploit variation in the rollout of INPRES across Indonesian districts.

This enables us to utilize two sources of variation: 1) geographic variation from the intensity

of primary schools built across districts; and 2) cohort variation by comparing individuals

who were of primary school age or younger to those who were older than primary school age

1Other recent papers that use the INPRES project include: Ashraf et al. (2018) who study bride priceand female education; Bharati et al. (2016) who analyze program impacts on adult time preferences; Bharatiet al. (2018) who examine whether INPRES mitigates the effects of adverse weather shocks; and Karachiwallaand Palloni (2019) who examine the impacts on participation in agriculture.

3

at the inception of the program.

As far as we are aware, we are the first paper to use longitudinal data from the Indonesian

Family Life Survey (IFLS) and the Indonesian Family Life Survey-East (IFLS-E) to explore

the long-term health and intergenerational effects of INPRES. The longitudinal aspect of the

IFLS allows us to observe household members over time, including those who have split-off

from the original household and formed new households. We find that fewer than half of

all children born to parents potentially exposed to INPRES are still coresident with their

parents during adulthood and therefore would be missing in analyses that only use cross-

sectional data. Furthermore, recent studies have shown that relying on coresident samples

can lead to biased estimates of intergenerational mobility (Asher et al., 2020; Emran and

Shilpi, 2018). In addition to providing new estimates from a representative intergenerational

sample, the rich set of questions in IFLS allows us to track several useful markers of health

for both generations, as well as a range of human capital outcomes for the second generation.

Our first set of results document that there were significant and meaningful effects on the

long-run health of individuals in the first generation. We create a summary index of several

health measures: self-reported poor health, the number of days missed daily activities due

to health reasons, chronic conditions, and depressive symptoms. About 40 years after the

program, we find that health is improved by 0.04 to 0.06 standard deviations (SDs) per

school built per 1,000 children relative to the comparison group. We also examine gender

differences and find that these effects are generally stronger for women.

Our second set of results examine the impacts on the children of cohorts exposed to the

rollout of INPRES, i.e. the second generation. Specifically, maternal access to INPRES

elementary schools increases children’s height-for-age by 0.06 SDs and reduces the likelihood

of childhood stunting by 7% of the mean. We also find improvements in children’s self-

reported health. We construct an index of health outcomes for the second generation and

find an improvement in health of about 0.03 SDs. We also examine children’s test scores

in the national 9th grade examination and find that children born to women exposed to

4

the INPRES program score between 0.08 and 0.10 SDs higher. In general, we find similar

impacts for sons and daughters, and smaller and statistically insignificant effects from father’s

exposure to the INPRES program. Overall, our results for the first and second generations

are robust to alternative specifications and we show evidence from event study analyses and

placebo regressions on the comparison group that validate the empirical strategy.

There are several pathways through which improved human capital (due to INPRES)

could lead to better first and second generation outcomes. These include assortative mating,

fertility, household resources, neighborhood quality and migration.2 We find mixed evidence

which suggests that women exposed to INPRES are more likely to marry better educated

men. With respect to fertility, we find no evidence that INPRES changed the timing of

the first birth or the spacing of births. While we find that women exposed to the program

experience a decline in their total fertility, our “back-of-the envelope” calculations imply

that reduced fertility only explains up to 8 percent of our second generation effects. We also

show that individuals exposed to the program have greater household resources as measured

by per capita consumption and housing quality. In addition, we observe that individuals

exposed to INPRES are more likely to live in communities with better access to health

services. Lastly, our evidence suggests that migration responses do not drive our results.

We also assess the success of the INPRES intervention by conducting a cost-benefit

analysis that accounts for the cost of construction and maintenance of the schools.3 A key

conclusion of our analysis is that including the spillover gains generated by the program, as

well as the effects on both generations makes a very big difference, raising the internal rate

of return from 8 percent to as high as 24 percent. Thus, traditional cost-benefit analyses

of school building interventions that only account for the labor market returns to education

(e.g. Aaronson and Mazumder, 2011; Duflo, 2001) may be far too conservative in their

assumptions and may significantly underestimate the full societal benefits.

2One important potential channel is parental investments in human capital. However, we do not havevery good measures to proxy for such investments. In addition, general equilibrium effects may also influenceboth the first and second generations effects (Duflo, 2004).

3See Appendix C for the details of our calculations.

5

Our paper contributes to several literatures. First, a vast literature in economics has doc-

umented the long-term effects of different types of social interventions on multiple dimensions

of human capital outcomes, mainly in high-income countries (see Almond et al., 2018 for a

review of this evidence). There is also emerging evidence from lower income countries on

the long-term effects of interventions, which mainly analyzes demand-side educational in-

terventions such as conditional cash transfers (Parker and Vogl, 2018), vouchers (Bettinger

et al., 2017) and compulsory schooling laws (Aguero and Ramachandran, 2018). We add

to this literature by examining the long-term effects of a common supply-side educational

intervention in the form of school construction that aims to improve schooling access.

Second, we provide new evidence on the causal effects of schooling on health in the context

of a middle income country. Most of the evidence thus far relies on randomized variation in

pre-school access or quasi-experiments that exploit changes to compulsory schooling laws.

Thus far, the findings on the causal effects of education on health (or mortality) are quite

mixed (see e.g. Galama et al., 2018; Oreopoulos and Salvanes, 2011 for a review).4

Third, our paper contributes to the emerging literature on the intergenerational effects of

social interventions. Relatively few studies have been able to causally identify the intergen-

erational effects of policies. This is mainly due to the demanding data requirements needed

to answer this question. In particular, the analysis calls for data that measure outcomes

in two distinct generations and link family members over time. Most existing studies on

intergenerational effects have been done in the context of high-income countries.5 There

is limited evidence in low and middle income countries (Aguero and Ramachandran, 2018;

Grepin and Bharadwaj, 2015; Wantchekon et al., 2014) and we are among the first to study

the intergenerational human capital impacts of a national school construction program in a

4Plausible mechanisms have been proposed for how education can improve health. One theoretical per-spective hypothesizes that education can improve productive efficiency and allocative efficiency (Grossman,1972; Kenkel, 1991). Through productive efficiency, more education leads to a higher marginal product fora given set of health inputs. Through allocative efficiency, higher educated individuals choose more efficientinputs into health investment. Examples of these efficiencies include greater financial resources, improvedknowledge, better decision-making ability and changes to time preferences.

5For example, evidence from Head Start and Medicaid in the US (Barr and Gibbs, 2018; East et al.,2017), and compulsory education in Sweden and Taiwan (Chou et al., 2010; Lundborg et al., 2014).

6

relatively lower income setting. The evidence of whether program effects persist and trans-

mit to the next generation is highly policy relevant as current debates about the funding for

such social programs may underestimate their benefits.

The remainder of the paper is organized as follows. Section 2 provides an overview of

the program, Section 3 describes the data. Section 4 describes the methods used to estimate

the program effects on the first generation and the long-term effects. Section 5 describes

the methods and results on the second generation. Section 6 presents additional robustness

checks, and 7 discusses some potential mechanisms. Section 8 concludes with cost-benefit

calculations and policy implications.

2 Background

2.1 The Indonesian context and INPRES Program

Primary school enrollment in Indonesia was around 65% in the 1960s (Booth, 1998) and

the illiteracy rate in the 1971 Census was 31%. In 1973, President Suharto took advantage

of the oil boom and created the INPRES Primary School program (Sekolah Dasar INPRES,

or SD INPRES) to expand primary school access and other INPRES programs to promote

regional economic development (Duflo, 2001). The SD INPRES program, which began its

implementation in 1974, sought to reach at least 85% of primary school aged children, who

usually start their six years of primary schooling between the ages of 6 and 7. To increase

equity in access to basic education, the SD INPRES program targeted places with low

primary school enrollment (Duflo, 2001). Provinces outside the main island of Java, which

have been traditionally poorer and more rural, received more funding to allow for the building

of more schools in those areas. By 1979, the SD INPRES program had constructed over

61,000 primary schools, increasing the number of schools available by 2 schools per 1,000

children (Duflo, 2001). This rapid growth makes this program one of the fastest primary

school construction interventions and a successful case of a large-scale school expansion on

7

record (World Bank, 1989).

INPRES schools were provided with school equipment and adequate water and sanitation

(Duflo, 1999; World Bank, 1989). These new schools also included improvements in classroom

spacing, thereby avoiding the double shifts of students (where some students would attend

the morning session while others would attend the afternoon session, resulting in a shorter

instruction time per student). These schools were also staffed with teachers to attain a

ratio of one teacher per 40 to 50 students, which was an improvement from a ratio of 50

to 60 students per teacher in the 1950s.6 The school construction was accompanied by

improvements in teacher training through a program to construct primary teacher training

schools (World Bank, 1989) and increases in teacher salaries. The SD INPRES program

efforts to increase primary school completion were accompanied by the elimination of primary

school registration fees in 1977, all of which contributed to an increase in primary school

enrollment to 91% by 1981.

Another INPRES program focused on improving water and sanitation. This program

sought to build 10,500 piped water connections and 150,000 toilets in villages. This program

was distributed based on the pre-program incidence of cholera and other diarrheal diseases,

access to clean water, the availability of hygiene and sanitation workers, and a preliminary

survey. Following Duflo (2001), we include the water and sanitation program as a control

variable in our analysis of the SD INPRES program to avoid confounding issues.

2.2 Previous evidence on school construction interventions

Previous evidence on INPRES

A handful of studies beginning with Duflo (2001) have examined the impacts of the

INPRES primary school construction program on outcomes using a difference-in-difference

approach that exploits the geographic intensity in the construction of schools and cohort

variation. Duflo (2001) focused on men born between 1950 and 1972 to examine the effects

6https://unesdoc.unesco.org/ark:/48223/pf0000014169eng. Last accessed May 22, 2019.

8

of the program on educational attainment and wages. She finds that an additional school

built per 1000 school-aged children increased years of schooling by 0.12 to 0.19 years and

wages by 1.5 to 2.7 percent. Duflo (2004) also studies the general equilibrium effects of this

large program and shows that the increase in education among exposed individuals increased

their participation in the formal labor market and had a negative effect on the wages of older

cohorts.

Breierova and Duflo (2004) study the effects of mother’s and father’s education on child

health and find that parental education reduced infant and child mortality. Martinez-Bravo

(2017) examines the effect of the INPRES school construction program on local governance

and public good provision and finds that the program led to a significant increase in the

provision of public goods, such as the number of doctors, the presence of primary health

care centers, and access to water. A potential mechanism behind these effects is the increase

in the education of the village head.

A contemporaneous paper to ours by Akresh et al. (2018) examines the long-term effects

of INPRES on the socio-economic well-being of the first generation and the intergenerational

effects on school attainment. Their analysis is complementary to ours, studying a different set

of outcomes and using nationally representative cross-sectional data from the 2016 Susenas,

which relies on adult children co-resident with their parents. We discuss potential concerns

to using cross-sectional data for intergenerational analysis in the data section (Section 3).

We build on these studies that document positive effects on the first generation exposed

to INPRES to examine whether these individuals have better health 40 years after the

intervention and whether these gains transmit to the next generation’s human capital.7

7Other recent papers that use the INPRES project include: Ashraf et al. (2018) who study bride priceand female education; Bharati et al. (2016) who analyze impacts on adult time preferences; Bharati et al.(2018) who examine whether INPRES mitigates the effects of weather shocks; and Karachiwalla and Palloni(2019) who examine the impacts on agriculture.

9

Other school construction interventions

Improving access to education through school building is a popular supply-side interven-

tion in low and middle income countries, where students may need to travel long distances to

reach the closest school (Glewwe and Kremer, 2006; Kazianga et al., 2013). Several studies

have examined whether improvements in school infrastructure have causal effects on enroll-

ment and various short and medium term student outcomes such as test scores (for a recent

review of the literature, see Glewwe and Muralidharan (2016)). However, these studies did

not examine longer-term outcomes, intergenerational effects or other dimensions of human

capital such as health. A few studies of historical interventions have begun to explore such

outcomes. These include Wantchekon et al. (2014) who found that colonial era missionary

schools built in Benin had human capital spillovers as well as intergenerational effects, and

Chou et al. (2010) who found that the 1968 Taiwanese compulsory schooling reform (which

included a school construction component) improved birth weight and infant health in the

next generation.

A related literature has analyzed the effects of the Rosenwald Schools which were built

for blacks living in rural parts of the American South largely during the 1920s. The insti-

tutional setting for these schools is similar to what many developing countries face today.

Aaronson and Mazumder (2011) show that the schools led to significantly higher educational

attainment and test scores. Aaronson et al. (2014) show that the Rosenwald schools impacted

fertility patterns and Aaronson et al. (2020) find that the schools improved long-term health.

2.3 The effect of education on health

One of the most striking findings in the social sciences is the gradient in health and

mortality by socioeconomic status (e.g Cutler et al., 2011). However, whether this association

is truly causal and whether it can be influenced by educational polices is less clear and

may depend on the context. If large-scale school building programs such as INPRES can

improve not only educational attainment but also provide meaningful spillovers to health,

10

then the case for such interventions become even more salient. Furthermore, if educational

improvements also result in health improvements into the next generation, then the case for

these policies is even stronger.

The empirical evidence on the causal effects of education on health appears to be mixed

but few studies have specifically explored the effects of school construction in the context

of a lower income country (see Galama et al. (2018) for a recent review).8 One possible

explanation for the mixed findings in the literature is that education may induce other

behavioral changes such as migration, and depending on the setting, these other behavioral

changes can lead to worse health and therefore obscure the health promoting aspects of

education (Aaronson et al., 2020). Unlike much of the previous literature that has examined

pre-school or compulsory schooling reforms, we provide evidence on the effect of expanding

access to primary education on an individual’s own physical and mental health in a lower

income setting.

A growing number of studies extend the analysis of the effects of education on the health of

offspring in the next generation. In general, parental education, especially that of the mother

has been shown to be a strong predictor of children’s outcomes, such as birth weight (Currie

and Moretti, 2003). However, changes in compulsory schooling laws in several countries

have resulted in mixed evidence. Evidence from the UK shows little effects on child health

(Lindeboom et al., 2009), while positive effects have been found in Taiwan (Chou et al., 2010),

Zimbabwe (Grepin and Bharadwaj, 2015) and Turkey (Dursun et al., 2017). Previous work

in our setting by Breierova and Duflo (2004) found that INPRES led to lower child mortality.

While these studies focus on children’s outcomes before the age of 5, we examine children’s

outcomes when they are older to evaluate the persistence of the health effects. Our study

contributes to the small but growing literature on the intergenerational effects of a large-scale

8Studies that exploit compulsory schooling laws across various countries (e.g. Clark and Royer, 2013,Oreopoulos, 2007) affecting different age groups in different time periods provide mixed evidence on theeffects of education on various health outcomes (Mazumder, 2012). For smoking and obesity, analyses fromseveral high-income countries shows mixed evidence and some heterogeneity by gender and demographiccharacteristics (Galama et al., 2018).

11

education program in lower income countries.

3 Data

We use data from the Indonesian Family Life Survey (IFLS). Our IFLS sample combines

the main IFLS and the IFLS-East (IFLS-E). The main IFLS is a longitudinal household

survey that is representative of approximately 83 percent of the Indonesian population in

1993. Subsequent waves (IFLS 2 to 5) in 1997, 2000, 2007 and 2014 sought to re-interview

all original households, as well as any households that had split-off. The IFLS-E, conducted

in 2012, is modeled after the main IFLS and covers 7 provinces in the eastern part of

Indonesia that were excluded by the main IFLS.9 In 1993, IFLS-1 included 7,224 households

residing in 13 provinces, which covered more than 200 districts. The IFLS-E included 2,500

households residing in seven provinces in eastern Indonesia, which covered about 50 districts.

Thus, the main IFLS and IFLS-E encompassed almost 300 of Indonesia’s 514 districts. The

IFLS is well suited for our analysis since the main IFLS is longitudinal, allowing us to not

only track long-term outcomes of adults in the original survey but to also follow children into

adulthood as they form new households. As we show below, this is important for maintaining

a representative sample of the families that were potentially affected by INPRES. The IFLS

is also valuable as it includes a comprehensive set of socio-demographic measures of interest.

Sample of interest We analyze the long-term and intergenerational outcomes of first

generation individuals who were born between 1950 and 1972. Those born between 1950

and 1962 were older than primary school age (age 12) at the time of INPRES (in 1974)

and thus were not exposed to the new schools, while those born between 1963 and 1972

were younger than age 11 during INPRES and thus could benefit from the primary school

expansion. We call this sample the “expanded sample”. It is worth nothing that the treated

9In a companion paper, we compare the estimated effect of the INPRES program on primary schoolcompletion using the nationally representative Intercensal Census (SUPAS) and the SUPAS restricted to theIFLS and IFLS-E provinces and find similar estimates (Mazumder et al., 2019).

12

group in this sample comprises individuals partially and fully exposed to the INPRES schools.

Those partially exposed were older than age 7 but younger than age 12, so only a part of their

primary school ages occurred after the program onset, while those who are fully exposed were

younger than age 7 at the time of the program, thus exposed to INPRES schools during all

their primary school years. Following Duflo (2001), we also present estimates for a sample

that defines the treated group as those who were fully exposed (born between 1968 and

1972) and the non-exposed group as individuals who are closer in age (born between 1957

and 1962). We call this sample the “restricted sample”. In both samples, individuals have

completed most of their fertility cycle by 2012 or 2014, which enable us to examine their

children’s outcomes.

We believe it is useful to consider results for both the expanded and restricted samples.

The main advantage of the expanded sample is that it enables us to have greater statistical

power. While the restricted sample has less statistical power, it uses cohorts where the

treatment was more targeted.

Long-term Outcomes Our outcomes of interest include measures of self-reported health.

Adult respondents were asked to report their physical health through a series of questions on

self-reported health status and chronic conditions. Using this information, we construct the

following outcomes of interest: good self-reported health, the number of days a respondent

missed his or her activities due to health reasons in the past 4 weeks prior to the survey,

any diagnosed chronic conditions, and the number of diagnosed conditions. Some of these

self-reported health measures have been found to be highly predictive of well-known health

markers such as mortality in both high and lower income countries (DeSalvo et al., 2005;

Idler and Benyamini, 1997; Razzaque et al., 2014).

We also examine mental health outcomes based on self-reported depressive symptoms.

The IFLS uses 10 questions from the Center for Epidemiologic Studies Depression Scale

(CES-D), which has been clinically validated. We score each symptom in the 10 questions

13

and use the sum of the scores based on reported symptoms, where higher scores indicate a

higher likelihood of having depression.10 Because some of our outcomes of interest are only

assessed for individuals older than 40 years old, we use information from the IFLS-E in 2012

and IFLS-5 in 2014.

To summarize the multiple health outcomes, we construct a summary index following

Kling et al. (2007) and Hoynes et al. (2016). The index also addresses concerns due to mul-

tiple hypothesis testing. We standardize each health outcome by subtracting the mean and

dividing by the standard deviation of the comparison group and equalize signs across out-

comes, so that higher values of the standardized outcomes represent poorer health outcomes.

Then, we create a summary index variable that is the simple average of all standardized out-

comes. The components include self-reported poor health (instead of self-reported good

health), the number of days missing one’s primary activity due to health reasons, any di-

agnosed chronic conditions (e.g. hypertension, diabetes), the number of chronic conditions,

and mental health score.11 This summary index is our main outcome of interest and we also

present results for each component.

In addition, we separately explore two other health outcomes but do not include them

in our summary index: Body Mass Index (BMI) and high blood pressure. It is not clear

apriori whether BMI is a positive or negative outcome since there is evidence of a positive

association between SES and BMI in developing countries, including Indonesia (Dinsa et al.,

2012). High blood pressure (systolic pressure higher than 130 mm Hg or diastolic pressure

higher than 80 mm) is only measured on the survey date, so no diagnosis of hypertension

is made. We prefer to use self-reported diagnosed hypertension, which is one component of

the chronic conditions included in the summary index discussed above.

10Details of the construction of these variables are available in Appendix A.11We exclude health care utilization from our analysis due to the low incidence of preventive care in

developing countries, including Indonesia. In our sample, only 8% of respondents reported obtaining at leastone general health check-up in the 5 years prior to the survey.

14

Intergenerational Outcomes Respondents born between 1950 and 1972, the first gen-

eration, have children who were born between 1975 and 2006, which make up the second

generation. Regarding children’s health, the IFLS collects the following health measures:

height and weight, hemoglobin count, and self-reported general health status (which was

obtained from the primary caregiver for respondents under 15). Given the timing of the

IFLS survey years and the wide range of birth years of the second generation, the oldest

individuals in the second generation were 18 in 1993 (in IFLS-1) and the youngest were 8

in 2014 (in IFLS-5). Therefore, for the second generation’s health outcomes, we focus on

children aged between 8 and 18 in each wave of the IFLS.

Our second generation’s health outcomes of interest include height-for-age and stunting,

defined as more than two standard deviations below the mean in height-for-age z-score.12

These health measures capture children’s general growth trajectory and proxy for long-term

and cumulative nutritional status. We also include anemia from children’s hemoglobin count

as another proxy for children’s nutritional status and self-reported general health to capture

overall health status. Similar to the first generation long-term outcomes, we summarize

these health outcomes for the second generation by creating a health index. The components

include being stunted, being anemic, and having self-reported poor health (instead of self-

reported good health), thus higher values represent poor health outcomes.13

Additionally, the IFLS includes children’s educational history, which allows us to obtain

children’s primary and secondary school completion as well as scores on the national primary

school (6th grade) and secondary school (9th grade) examinations.14 For consistency of

comparison across the years, we create z-scores for each year of examination.15 In this study,

12We use the 2007 WHO growth chart, which is applicable to children between 0 and 19 years of age.13Details of the construction of these variables are available in Appendix A.14While Indonesia has had national examinations since 1950s, the standardized national curriculum was

only implemented in 1975 and the national examination after the curriculum standardization began in 1980.Thus, test scores are only available for the second generation. Indonesia implemented 6 years of compulsoryschooling in 1984 and expanded that to 9 years of compulsory schooling in 1989, which corresponds to lowersecondary. We are not able to examine the end of secondary school, which is 12th grade, examinations sincemost of the children are too young in 2012 and 2014.

15The scoring of the national examination changed between the 1980s and 2000s.

15

we focus on the secondary school examinations since our companion paper, Mazumder et al.

(2019), analyzes the intergenerational impacts of INPRES on primary test scores.

INPRES Exposure variable We combine administrative data on the number of INPRES

primary schools built between 1974-1979 at the district level with the IFLS.16 We assign ge-

ographic exposure to the INPRES program using the number of schools constructed in the

first generation’s district of birth, since using district of residence during respondents’ pri-

mary school age is potentially endogenous. For the second generation, we use the household

roster to identify biological mother-child and father-child pairs born to the first generation

individuals, who were born between 1950 and 1972.17

Summary Statistics The first generation sample includes around 12,000 adults born

between 1950 and 1972 with information on their place of birth and observed at any wave

of the IFLS.18 For the first generation long-term outcomes, the analyzed sample consists of

around 10,200 individuals observed in the IFLS-E in 2012 or the IFLS-5 in 2014 (Table A.1,

Panel A). The average number of INPRES schools built in the first generation’s district of

birth is 2.1, consistent with the national average. The first generation sample is balanced

across gender. About 60% of the sample were born between 1963 and 1972, and 46% of the

sample are Javanese, the main ethnic group in Indonesia. First generation individuals score

an average of 5.5 on the mental health screening questionnaire, 72% of adults in the sample

reported being healthy and 42% reported having at least one diagnosed chronic condition.

16Appendix Figures A.1 and A.2 show the comparison between the national INPRES rollout and theprovinces covered by the main IFLS and IFLS-E. We are not aware of other previous studies that haveutilized the IFLS-E to study the effects of INPRES on these outcomes.

17Details of the construction of these variables are available in Appendix A.18A potential drawback of the IFLS data is its small sample size. To address this, we performed power

calculations that validate the use of this data. Based on the number of birth districts (clusters) in theIFLS, which corresponds to 333, a control group with a mean of 0 and a standard deviation of 1 andan intraclass correlation of 0.03 (which is the ICC for the health indices), the sample of first generationindividuals in the main IFLS and the IFLS-E allows us to detect treatment effect sizes between 0.03 and0.04 standard deviations at the 10% and 5% significance levels with 80% power. For the first-order outcomeof the intervention, primary school completion, the IFLS sample can detect effect sizes between 0.03 and0.04 percentage points at the 10% and 5% significance level with 80% power.

16

The second generation sample consists of about 10,000 individuals who are the children

of the first generation sample (Table A.1, Panel B). The sample is balanced across gender,

and about 40% are first-born. The average year of birth is 1988. About half of the children

have mothers who were born between 1963 and 1972, and similarly, about half have fathers

who were born between 1963 and 1972. Children’s health measures were taken in each wave,

and we use observations between the ages of 8 and 18. The average height-for-age z-score

is -1.6 and 36% are stunted. About one fourth of the second generation have anemia, and

almost 90% have good self-reported health.

The Benefit of Longitudinal Data for Intergenerational Analysis A key advan-

tage of our IFLS data is that for the vast majority of our sample, we are able to follow

split-off households that are formed after children enter adulthood and leave their parents’

household.19 Therefore, our intergenerational sample is largely representative of the universe

of children born to parents who could have been exposed to INPRES. In contrast, studies

in developing countries have often relied on cross-sectional data to form intergenerational

samples, thereby selecting on children who are still co-resident with their parents in adult-

hood.20 Recent studies have shown that relying on coresidency can lead to biased estimates

of intergenerational mobility (Asher et al., 2020; Emran and Shilpi, 2018).

We highlight the potential bias that would have occurred had we relied on purely cross-

sectional data by showing how the representativeness of the sample is altered if we were to

only use coresident children in 2014, which is the latest year in our data. In our second

generation sample, we find that around 55 percent of children born to the first generation

sample are no longer coresident with their parents by 2014 in IFLS-5. This highlights the

19For a small subset of our intergenerational sample (less than 10%) that is formed using the IFLS-E,we only include children who were co-resident with their parents as adults. Attrition rates across the fivewaves of the IFLS are relatively low: the original household re-contact rate was 92% in IFLS-5, and 87.8%of original households in IFLS-1 were either interviewed in all 5 waves or died (Strauss et al., 2016).

20In a study closely related to ours, Akresh et al. (2018) estimate intergenerational effects of INPRES usingcross-sectional data from 2016, where parent-child pairs require co-residence. Recent examples of studiesin developing countries that have formed intergenerational samples based on coresidence include Nimubonaand Vencatachellum (2007) and Emran and Shilpi (2015).

17

fact that in our context, relying on coresidency of adult children with their parents leads to

a dramatic loss of sample.

We compare the characteristics of co-resident and non co-resident children in IFLS-5

(Table A.2). Co-resident children are more likely to be younger, come from households with

a higher asset index21, their parents have higher education, and their mothers were older at

the time of their first birth. We also compare the characteristics of co-resident children in

IFLS-5 to co-resident children in a nationally representative survey, the 2014 Socioeconomic

survey (2014 Susenas) to show that coresident families in the IFLS are similar to the national

sample (Table A.3). When we restrict the 2014 Susenas to all IFLS provinces, the co-resident

child characteristics are similar to the national survey sample (Table A.3, columns 4-6).22

Overall, the IFLS data provide some clear advantages over cross-sectional samples when

studying the intergenerational effects of policies by ensuring a more representative sample.

Our analysis suggests that studies that use cross-sectional data and select on adult children

who are coresident with their parents rely on fewer than half of the intended universe of

families and tend to be positively selected. The main disadvantage of a survey like the IFLS,

of course is the relatively small sample size, which limits statistical power. Therefore, given

the tradeoff between sample size and bias, it may be useful to consider results from both the

IFLS as well as results from other studies that use larger samples but may suffer from bias

due to selection on coresidence.

21The asset index includes the ownership of the following assets: savings, vehicle, land, TV, appliances,refrigerator, and house.

22Additionally, when the 2014 Susenas is restricted to the main IFLS provinces, the co-resident childcharacteristics are similar to the co-resident children in IFLS-5 (Table A.3, columns 7-9).

18

4 The First Generation

4.1 Estimation Strategy

In the first part of the analysis, we estimate the long-term effects of INPRES on outcomes

measured roughly forty years later. Following Duflo (2001), we exploit variation in exposure

to the primary schools by birth cohort and geography as described in Section 3. We estimate

the intent-to-treat effects using the following equation:

yidt = β(exposedt × INPRESd) +∑t

(Pd × τt)δt +Xidtγ + αd + τt + εidt (1)

where yidt is the outcome of interest for individual i born in district d in year t. exposedt is a

dummy variable equal to 1 if individual i was born in the relevant birth cohorts exposed to

INPRES. In the expanded sample, this indicator takes the value of 1 for cohorts born between

1963 and 1972, while in the restricted sample, the exposed cohorts were born between 1968

and 1972. INPRESd captures the intensity of the program: the number of schools (per

1,000 school-aged children) built in birth district d during the school construction program.

αd and τt are district and year-of-birth fixed effects. Pd × τt captures birth-year fixed effects

interacted with the following district-level covariates: the number of school-aged children in

the district in 1971 (before the start of the program), the enrollment rate of the district in

1971 and the exposure of the district to another INPRES program: a water and sanitation

program. These interactions control for the factors underlying the allocation of the INPRES

schools and for other programs that could confound the INPRES school effects. Xidt is a

set of individual characteristics: gender, ethnicity (Javanese indicator) and month of birth

fixed effects.23 Standard errors are clustered at the district of birth level. β measures the

23Our specification modestly improves upon Duflo (2001) in two ways. We use an ethnicity dummy forwhether the individual is Javanese and month of birth dummies. We control for being Javanese since theyare the largest ethnic group in Indonesia and in our sample, and the group has different means. We includemonth of birth to control for potential seasonality (Yamauchi, 2012). In Mazumder et al. (2019), we alsoestimate a model for primary school completion that uses the identical covariates as Duflo (2001) by excludingmonth of birth and ethnicity, and find similar effects.

19

effect of one school built per 1,000 children. We estimate the models pooled and separately

by gender.

The causal effect of exposure to INPRES schools is identified as long as the program

placement of schools across districts is exogenous conditional on district of birth, cohort fixed-

effects, and the interactions of year of birth and district level covariates. Hence, if before the

construction of schools in 1974, districts with high program intensity had differential growth

in educational outcomes compared to low program intensity districts, this might suggest

that our identification assumption is not credible. Using the IFLS raw data, we show the

similar trends in primary school completion for cohorts who finished primary school before

the program was implemented in high and low program intensity districts (Figure B.1).24

To validate our empirical strategy, we also estimate equation 1 for pre-program cohorts to

examine pre-trends for the health and education outcomes of interest in the first generation.

We find no significant pre-trends for cohorts before the INPRES school program (Figures B.2

and B.3). Additionally, we perform placebo regressions using the IFLS data on comparison

cohorts (individuals born between 1950 and 1962) for each health and education outcome

(Figure B.4). These regressions define a “placebo exposed group” as cohorts born between

1957 and 1962, while those born between 1950 and 1956 serve as the comparison group.

We find small and statistically insignificant effects on all of our health outcomes, thereby

providing reassuring evidence on the absence of pre-trends before program implementation.

4.2 First Generation Long-term Effects

In our companion paper, Mazumder et al. (2019), we show evidence of the first order ef-

fects of the INPRES school construction program on primary school completion, the margin

that the program targeted. We replicate the estimates in Table B.1, which has two panels:

Panel A presents estimates for the “expanded sample”, comprising of individuals born be-

tween 1950 and 1972. The treated cohorts, those born between 1963 and 1972, pool partially

24Following Duflo (2001), high program intensity districts are defined as districts “where the residual of aregression of the number of schools on the number of children is positive”.

20

and fully exposed individuals. Panel B presents estimates for the “restricted sample” that

excludes the partially exposed. This sample comprises of individuals born between 1968-

1972 in the treated and those born between 1957-1962 in the comparison group. Each panel

has three columns: all, male and female. The results suggest that exposure to the INPRES

primary schools increased the probability of completing primary school for both men and

women. The effects range between 2.5 and 3 percentage points per primary school per 1,000

children in the “expanded sample”, while the effects are between 3 and 5 percentage points

among individuals fully exposed to the program in the “restricted sample”.25

We next turn to analyze the long-run health outcomes in the first generation. We begin by

examining the poor health index, which summarizes the following outcomes: self-reported

poor health (indicator that takes the value one if a respondent reported his or her self

status as not “very healthy” or “not healthy”), the number of days missing one’s primary

activity due to health reasons, any chronic conditions, the number of chronic conditions,

and mental health score. These outcomes are measured in 2012 (IFLS-E) and 2014 (IFLS-5)

and respondents in the sample were at least 40 years old when they were surveyed. Figure

1 and Table 1 (col. 1) show that adults exposed to the INPRES primary school experience

a decline in the poor health index by 0.04 standard deviations for each additional primary

school constructed (per 1,000 children).

We also perform an event study analysis (Figure 2). To reduce the noisiness in the data,

we create bins, with each bin containing three years of non-overlapping birth cohorts. We plot

the relationship between the poor health index and the first generation’s age when INPRES

was implemented by estimating an equation similar to equation 1. For cohorts who were too

old to be exposed to the program, we find no significant pre-trends among these cohorts as

25To compare our results to a nationally representative sample, we also use the 2014 SocioeconomicSurvey (2014 Susenas) to examine the program effect on primary completion (Table A.4). The 2014 Susenasis representative at the district level. Using the 2014 Susenas, the program effect on primary completion forall provinces ranges from 1.1 to 1.7 percentage points. When we restrict the Susenas to the IFLS and IFLS-Eprovinces, the program effect on primary completion ranges from 1.5 to 2.5 percentage points, similar to ourfindings using the IFLS and IFLS-E. These results provide evidence of the representativeness of the IFLSand IFLS-E.

21

the coefficients bounce around zero. For exposed cohorts, the point estimate for each cohort

group is negative. While the estimated coefficients using the event study framework are not

statistically significant at conventional levels, they are consistent with our earlier estimate

using the difference-in-differences framework (shown in the far right of Figure 2). These

results show that individuals exposed to the INPRES program have improved overall health

about 40 years later.

We then analyze the components of the health index to examine each individual outcome

separately (Table 1, cols. 2-6). Using the expanded sample (Panel A), we find that an

additional school per 1,000 children improves good self-reported health by 3.9 percentage

points, which is about 6% relative to the mean (col. 2). Also, exposure to INPRES decreased

the number of days of missed activities by 0.17 days, which is almost 10% of the mean (col.

3). We find that an additional primary school per 1,000 children lowers the likelihood of

reporting any chronic conditions by 2.5 percentage points (5% of the mean, col. 4). In terms

of the number of diagnosed conditions (col. 5), INPRES exposure is associated with a 0.05

fewer diagnosed chronic illnesses for each additional primary school constructed (per 1,000

children). When using the restricted sample, the INPRES effects on these health measures

are similar but somewhat noisier (Panel B). Finally, we examine the program effect on adult

mental health as an additional marker of well-being (col. 6). We use 10 items on depressive

symptoms from the CES-D scale and use the sum of the scores. We find that adults exposed

to INPRES are less likely to report symptoms of depression. This effect is stronger and

statistically significant in the restricted sample when we focus on individuals who were fully

exposed to the program (an effect of 5% of the mean, Panel B). Overall, these findings are

consistent with the health improvement shown by the health index.

To address any concerns about inference due to multiple hypothesis testing, we test for

the joint significance of the components of the index and find that the they are jointly

significant in both the expanded sample in Panel A and restricted sample in Panel B (with

p-values of 0.0003 and 0.045 respectively). We also adjust the standard errors for multiple

22

hypothesis following Simes (1986) and most of our results are robust to this adjustment

(Table B.2).

We also estimate the program effects separately for men and women (Table 2 and Figure

1). We find that the program impact is concentrated among treated women, for whom the

effect corresponds to -0.06 standard deviations in the poor health index, and the gender

difference is marginally significant for the restricted sample (cols. 1-2, Panel B). When we

analyze the components separately, we find that the point estimates are generally significant

for women. For self-reported health status, there is a statistically significant increase of 6.2

percentage points for women (col. 4) whereas for men, the point estimate is 1.7 percentage

points and not statistically significant at conventional levels (col. 3). We test for the gender

difference using the seemingly unrelated regression model and find a significant difference,

suggesting a stronger effect for women. For the other health outcomes, using the expanded

sample (Panel A), the program effects on the number of days of missed activities, any

chronic condition, and the number of chronic conditions are significant for women, but the

estimated gender differences are not statistically significant. We find similar effects using

the restricted sample for those outcomes, and we also find that the program effect on mental

health is significant for women who were fully exposed to the program (although the gender

difference is not significant). These results suggest that the health improvement is generally

stronger for women, which may be an important channel for the intergenerational effect of

the program.

We also examine additional health outcomes: hypertension and Body Mass Index (BMI)

to supplement our set of self-reported health measures (Table B.3). We find about a 3%

reduction in the probability of having high blood pressure among individuals who were fully

exposed (Panel B, cols. 1-3). This is consistent with almost a 10% reduction in the self-

reported diagnosed hypertension (cols. 4-6), which is concentrated among women. The gap

between the measured high blood pressure and diagnosed hypertension may be the result of

low utilization of preventive care and the fact that a one time measure of high blood pressure

23

is not a diagnosis. As an additional health marker, we also include BMI (cols. 7-9).26 We

find about 0.2 higher BMI, about 0.8% increase from the mean, among exposed individuals.

We also estimate the probability of being overweight, defined as BMI above 25 and find that

the results are driven by high BMI. This is consistent with the positive association between

SES and BMI in many developing countries including Indonesia Sohn (2017). Also, this

finding is in line with evidence from a mandatory schooling reform in Turkey (Dursun et al.,

2018).

Lastly, we also analyze the effects of INPRES on health behaviors. In particular, we

examine the program effect on tobacco use and alcohol consumption for men and teen preg-

nancy for women (Table B.4).27 We find no significant effect on ever smoking, but we find

a 4.5% increase in the probability of smoking at the time of the survey, which is consistent

with the SES gradient of smoking in Indonesia (Triyana et al., 2019). We find no signif-

icant effect on the intensive margin of daily cigarette consumption. Additionally, we find

an increase in alcohol expenditure among men. For women, we find no significant change

in teen pregnancy.28 These results on risky behavior for men, and not for women, may

partly explain the absence of significant long-term health benefits among men exposed to

the INPRES program.

Taken together, those exposed to INPRES have better self-reported physical and mental

health, which suggests the link between improved education and health.29 We contribute

to the literature on the non-pecuniary effects of improving education by providing evidence

from a lower middle income country. Our results are consistent with evidence from Bharati

26In spite of poorer outcomes for some of these additional health measures, including these outcomes inthe health index yields similar results.

27Indonesia has the highest smoking prevalence among men at 67%. Since Indonesia is predominantlyMuslim, alcohol consumption tends to be low. Additionally, teen pregnancy may be high due to childmarriage (around 10%, UNICEF). Nonetheless, we examine these health behaviors as potential channels.

28We use 17 and 18 as alternative age cutoffs for teen pregnancy and find similar results.29A recent paper by De Chaisemartin and d’Haultfoeuille (2018) proposes a fuzzy DID framework to

examine treatment effects when the treatment is not sharp as there is no group completely untreated. Weexplored this alternative approach, but it is not feasible in our context as it requires a categorical classificationof the treatment (for example, high versus low-intensity INPRES exposure). We found that we did not havesufficient power with our sample sizes to implement this Wald-DID approach.

24

et al. (2016), who find that INPRES increased patience in adulthood. Theoretically and

empirically, more patience is associated with better health investments (Fuchs, 1980), and

the effects of INPRES on time preferences constitute a potential mechanism for our findings

on long-term health.30 In addition, many of the health markers examined in this section

have been shown to be highly correlated with adult mortality (Idler and Benyamini, 1997).

Magnitudes

The average number of INPRES schools built was 1.98 per 1,000 children. This implies

that on average, exposure to INPRES primary schools increases the probability of being

“healthy” or “very healthy” by about 12%-20% of the mean. Similarly, at the average

exposure, we find a 10% to 14% reduction in reporting any chronic conditions and a 11% to

15% reduction in the number of reported chronic conditions. For comparison, our findings

are in line with previous studies that have used changes in compulsory schooling laws (CSL)

to estimate the impacts of education on similar self-reported health outcomes, even though

such studies mainly focus on higher-income countries. For example, Mazumder (2008) and

Oreopoulos (2007) find that an additional year of schooling from CSL in the US, UK and

Canada reduces the probability of being in fair or poor health by around 20% of the mean.

Additionally, we find that first generation individuals exposed to INPRES had fewer mental

health symptoms, by 10% to 18% of the mean at the average level of INPRES exposure.

This finding is similar to the effects of changes in CSL laws on well-being: the CSL law

in the UK is associated with an increase in overall life satisfaction by about 6% of the

mean (Oreopoulos, 2007), while the estimated effect of education reforms on the reduction

in depression in several European countries (Austria, Germany, Sweden, the Netherlands,

Italy, France and Denmark) is about 7% (Crespo et al., 2014). Overall, our findings on the

long-term effects of improving access to primary school on health outcomes are similar to

estimates from previous literature that document the link between education and health.

30We are unable to examine time preference because the IFLS-E does not include this module.

25

5 The Second Generation

5.1 Estimation Strategy

In this section, we examine the effects of parental exposure to INPRES on second gener-

ation human capital outcomes, exploiting the longitudinal nature of the IFLS.31 Specifically,

we estimate the following equation:

yidt = β(ParentExposedt × INPRESd) +∑t

(Pd × τt)δt +Xiγ + αd + τt + +εidt (2)

where yitd is the outcome of interest for child i whose mother/father was born in district

d in year t. The interaction ParentExposedt × INPRESd captures parental exposure to

INPRES based on parental district and year of birth. Xi is a set of child characteristics:

gender, birth order, and year and month of birth dummies. αd and τt are parent’s district and

year-of-birth fixed effects. The rest of the variables are as defined in equation 1. Standard

errors are clustered at the parent’s district of birth level. As in the previous section, we

consider the effects separately for the expanded sample and the restricted sample. The

coefficient β captures the effect of parental exposure to one INPRES school built per 1,000

(first generation) children.

Our empirical strategy relies on the assumption that parental INPRES program exposure

is uncorrelated with unobserved characteristics that vary across districts over time that also

may affect the second generation’s outcomes. To test the parallel trends assumption, we

estimate equation 2 for children born to pre-program cohorts and find no significant pre-

trends in human capital outcomes among second generation born to parents not exposed to

INPRES (Figure B.5). Additionally, we perform a falsification test that uses the children

of adults in the comparison group (adults born between 1950 and 1962). In this placebo

31We rely on co-resident children in the IFLS-E.

26

regression, we assume that adults born between 1957 and 1962 were exposed to the program.

We find no statistically significant difference in children’s educational and health outcomes

(Figure B.6). These results suggest similar pre-trends in the outcomes across districts among

children born to adults not exposed to the INPRES program.

We start by estimating the models separately for mother’s and father’s INPRES expo-

sure. These estimates provide the reduced form effects of INPRES exposure of any one

parent separately but does not account for the possibility that both parents could have been

exposed. One possible concern with that specification is that the comparison group may

be contaminated. This is because the program could affect the marriage market (Akresh

et al., 2018; Ashraf et al., 2018; Zha, 2019). For example, a father in the comparison group

(older than primary school age at INPRES rollout) could marry a woman in the treatment

group and thus be indirectly exposed to INPRES. Therefore, we present estimates from

specifications where we include both maternal and paternal exposure.32

5.2 Intergenerational effects

Health We now analyze the intergenerational effects of INPRES on several measures of

child’s health. For each outcome, we show three sets of estimations using the extended

and restricted sample: only mother’s exposure, only father’s exposure and both parents’

exposure. We begin by analyzing the poor health index (Table 3 and Figure 3), which

summarizes the following outcomes: stunting, self-reported poor health, and anemia status.

Given the timing of the IFLS and the fact that second generation children were born between

1975 and 2006, the health measures for these children are observed between ages 8 and 18

32When estimating specification for both parents exposure, including the full set of mother’s and father’sdistrict of birth and year of birth fixed effects as well as the interactions between these districts of birthfixed effects and districts’ covariates is very demanding for our small sample sizes. Also, when estimatingeach parent separately, we find larger effects from maternal than paternal exposure to INPRES (see below).Therefore, we estimate models that include mother’s and father’s exposure and the full set of covariates forthe mother. For the father’s covariates, instead of father’s district of birth and year of birth fixed effects,we use province of birth and two-year bins for year of birth fixed effects. Also, we include the interactionsbetween the father’s year of birth (in two-year bins) and the father’s district-level covariates. Standard errorsare clustered two-ways at the mother’s and father’s district of birth. In general, specifications using bothparents are much more demanding on our data and lead to less precise estimates.

27

across IFLS survey years. Also, because these outcomes are measured in each wave, we may

observe a child multiple times. To take into account the multiple observations per child,

we add wave fixed effects to equation 2. Table 3 column 1 presents the effects of maternal

exposure and shows that second generation children born to mothers exposed to INPRES

experience a decline in the poor health index by 0.03 standard deviations for each additional

primary school constructed per 1,000 children. Table 3 column 2 finds that paternal exposure

to INPRES also decreases the poor health index but the effects are smaller and statistically

insignificant. Table 3 column 3 present specifications that include both maternal and paternal

exposure. We corroborate our previous findings, second generation children born to mothers

exposed to INPRES experience a decline in the poor health index by 0.04 standard deviations

for each additional primary school constructed (per 1,000 children), while paternal exposure

has a smaller and insignificant effects.33 Also, we examine the effects by the child’s gender

and find that the impacts are similar across sons and daughters (Table B.5).

Additionally, we perform an event study analysis in Figure 4 (Panel A for maternal

exposure and Panel B for paternal exposure). To reduce the noisiness in the data, we combine

three parent birth cohorts into one group. These figures plot the relationship between the

second generation’s poor health index and each parent’s age when INPRES was implemented

and the dots represent the point estimates of the coefficients of each parental birth cohort

interacted with the number of INPRES schools, similar to equation 2. For mothers and

fathers who were too old to be exposed to INPRES primary schools, the point estimates

bounce around zero, which provide evidence of little pre-trends among non-exposed cohorts.

In contrast, we see that children born to mothers exposed to the program (who were primary

school age or younger) exhibit a decline in their poor health index. Some of these estimated

coefficients are noisy but they are consistent with our earlier estimate using the difference-

33The use of multiple observations per individual may be a concern since a child in the IFLS may bebetween ages 8 and 18 in several waves of the survey. To address this, we estimate equation 2 using theaverage outcome across waves and restricting the sample to one observation per child, weighted by thenumber of observations per child. The estimated effect is similar to our earlier findings. Children whosemothers were exposed to INPRES have 0.03 standard deviations lower average poor health index (TableB.8).

28

in-differences framework (shown in the far right). For fathers in the exposed cohorts, the

effects are smaller and noisier. Taken together, our findings suggest that maternal exposure

to a primary education intervention improves the health outcomes of their offspring.

Having established the index results, we next examine each of the components separately

(Table 4). We begin with two measures that capture the cumulative effects of health in-

vestments: children’s height-for-age (cols. 1-3) and stunting (cols. 4-6).34 We find a 0.06

standard deviation increase in children’s height-for-age z-score among those whose mothers

were exposed to the program (col. 1 and col. 3). We find small and statistically insignificant

effects among children whose fathers were exposed to the program (cols. 2 and 3). We

also find a reduction in stunting rates among children born to INPRES exposed mothers,

which is consistent with the estimated increase in height-for-age among these children (cols.

4-6 of Table 4). These children are 2 percentage points less likely to be stunted, which

corresponds to 7% of the mean. We find no significant effect through paternal exposure to

INPRES. Using the restricted sample yields qualitatively similar results. Given Indonesia’s

high stunting rates and the adverse effects of stunting, our results suggest that improved

access to education in one generation can spillover to enhance health in the next generation.

Turning to child’s self-reported health status, we show that maternal exposure to INPRES

increases the likelihood of being healthy by 1.2 percentage points (cols. 7-9). Finally, we

present the intergenerational effects on the child’s anemia status. While the point estimates

are negative for both mother’s and parent’s exposure, they are not statistically significant. In

general, when we focus on the restricted sample (Panel B), the estimated intergenerational

effects are noisy, but qualitatively similar. Overall, these results are consistent with earlier

work that finds an effect on infant mortality through maternal but not paternal exposure to

INPRES (Breierova and Duflo, 2004). To address multiple hypothesis testing, we test for

the joint significance of the components of the index. We find that for maternal exposure,

they are jointly significant in the expanded sample. We also adjust the standard errors

34We did not analyze weight indicators as those are considered more short-term health measures ratherthan a long-term cumulative indicator such as height.

29

for multiple hypothesis following Simes (1986) and most of our results are robust to this

adjustment (Table B.7).

Education We also examine whether parental exposure to INPRES resulted in improved

educational outcomes in the next generation. Previous evidence documents that children

born to women who were exposed to INPRES perform better on the national primary school

examination (Mazumder et al., 2019). We now extend the analysis to test scores on the

national 9th grade examination. We find that children of mothers who were exposed to

INPRES perform significantly better on the national secondary school examination (Table

5). On average, maternal exposure to one INPRES school (per 1,000 children) increases

their children’s secondary test scores by 0.08 standard deviations in the expanded sample

(Panel A) and 0.10 standard deviations in the restricted sample (Panel B). The estimated

effects are similar for sons and daughters in the expanded sample (Table B.6). Turning to

fathers’ INPRES exposure, the estimates are typically a little smaller and in no case are they

statistically significant (cols. 4-6). These findings are also consistent with our event study

analysis (Figure 4).

We also examine children’s lower secondary school completion. Here we find small and

statistically insignificant effects of maternal and paternal INPRES exposure (Table B.9,

cols. 1 and 2 respectively).35 The estimated effects of maternal and paternal exposure are

similarly small and not significant. Since the majority of the second generation individuals

in our sample are affected by Indonesia’s compulsory schooling laws, this could explain why

parental education has no effect on school completion but does appear to affect test scores.36

Taken together, mothers exposed to the program have children with better health and

educational outcomes, which suggests the importance of interventions that improve maternal

35The second generation was born between 1975 and 2006, therefore the majority should have completedprimary school by age 13 and the older children should have completed secondary school (9th grade) by theage of 16. We also examine children’s cognitive skills and find no significant effect.

36We did not examine high school completion since a significant fraction of second generation childrenare not old enough to be at that level of education. However, we will explore this as future IFLS becomeavailable.

30

education in order to improve children’s future outcomes. Our findings are consistent with

previous studies that have shown the intergenerational effects of social interventions in high

income countries on children’s health and education (Barr and Gibbs, 2018; Chou et al.,

2010; East et al., 2017; Lundborg et al., 2014; Rossin-Slater and Wust, 2018), and our

results contribute to the growing evidence in low and middle income countries (Aguero

and Ramachandran, 2018; Grepin and Bharadwaj, 2015; Wantchekon et al., 2014), where

intergenerational mobility tends to be lower than in high income countries.

Magnitudes

We now consider the size of the intergenerational effects on health and education. Under

INPRES, 1.98 schools were built per 1,000 students, which implies that, on average, maternal

exposure to INPRES raises height for age z-score (HAZ) by 0.12 SDs and reduces stunting

by 0.05 percentage points (14% of the mean). Ideally, we would like to compare these effect

sizes to the impacts of other social interventions in developing countries on similar intergen-

erational outcomes. However, such studies are nonexistent, highlighting the relevance of our

study. Therefore, for comparison, we refer to effect sizes from other interventions on similar

outcomes measured on those directly exposed in the first generation. For the case of Condi-

tional Cash Transfers, Fernald et al. (2008) find that a doubling of Mexico’s CCT program

led to a 0.16 SDs increase in HAZ and reduced stunting by 9% among children ages 10-14.

Similarly, Barham (2012) finds that exposure to the Matlab Maternal and Child Health and

Family Planning Program in Bangladesh increases HAZ by 0.2 SDs. Nores and Barnett

(2010) summarize the impacts of early childhood interventions in developing countries and

report that average effect sizes on long-term child health outcomes (ages 7 or above) are

around 0.12 SDs.

On average, maternal INPRES exposure increases secondary test scores between 0.16

and 0.20 SDs at the average level of INPRES exposure. In comparison, Baird et al. (2014,

2011) find that 2 years of exposure to a CCT program in Malawi that focused on 13-22 year-

31

old girls increased their English test scores between 0.13-0.14 SDs and math scores between

0.12-0.16 SDs. Similarly, Barham et al. (2013) examines the long-term effects of exposure

to the Nicaraguan CCT program in primary school on boys’ test scores ten years later and

find that the program increases average test scores by 0.2 SDs. For the case of merit-based

scholarship programs, Friedman et al. (2016) examine the effects of a scholarship program in

rural Kenya that targeted girls who were in grade 6 and find that their test scores increase

by 0.2 SDs in grades 10-11. Taken together, the effects of INPRES on second generation test

scores are comparable to the effect sizes of interventions like CCTs. Overall, the magnitudes

of our findings on the second generation are similar to estimates from studies that evaluate

other social interventions in developing countries.

6 Robustness

Alternative exposure variable Following Duflo (2001), our main specifications in equa-

tions (1) and (2) use the number of INPRES primary schools built between 1973-74 and

1978-79 per 1,000 children in the district of birth as a measure of the first generation’s ge-

ographic intensity of exposure to the program. This assumes that individuals are exposed

to the stock of schools at the end of the program. For robustness, we define an alternative

exposure variable using the number of schools constructed during an individual’s primary

school years (between the ages of 6 and 11) based on his/her age during the years of the

program implementation at his/her district of birth. Table B.10 shows our main results for

the first and second generation outcomes using this alternative specification and illustrates

that the estimated effects are similar in magnitude and statistical significance to those from

our main specification.

Alternative sample In addition, following Duflo (2001), our sample of interest corre-

sponds to first-generation men and women born between 1950 and 1972, which combines

non-exposed, partially, and fully exposed individuals. As an additional robustness check,

32

we expand our sample to include additional cohorts of fully exposed individuals: those born

up to 1975. These individuals are likely to have completed their fertility cycle by 2012-14

(IFLS-E in 2012 and IFLS-5 in 2014), thereby allowing us to observe the second generation’s

outcomes. Results using this alternative sample are remarkably similar to our main sample

estimates (Table B.11).

7 Potential mechanisms

In this section, we explore some of the possible mechanisms for our key findings. This

is a challenging endeavor both because there are a multitude of hypothesized channels and

because there are important data limitations. For example, ideally, we would like to be

able to explore the extent to which parents who were exposed to INPRES chose to invest

more in their children’s health through investments such as breastfeeding or vaccinations.

Unfortunately, we do not have data on these kinds of parental investments. Nevertheless,

there are several important channels which we are able to investigate using the richness of

the IFLS data. For each potential channel, we first present and discuss the effect of the

program on that channel. Second, when relevant, we perform some simple calculations to

assess the relative contribution of each mechanism in explaining the INPRES impact on

the first generation’s long-term health and the second generation’s health and education

outcomes.

In order to assess the potential contribution of each mechanism, we rely on relatively sim-

ple “back of the envelope” calculations. We do this by combining: i) estimated associations

between each mechanism and our outcome of interest in the comparison cohorts;37 ii) our

estimated effects of INPRES on each mechanism. We then compare the implied effects from

this exercise to our estimates of INPRES on the outcome of interest (See Table B.18). We

highlight two important limitations of this analysis. First, for part i), the associations we use

37The regressions to assess the associations include socio-demographic characteristics (i.e gender, ethnic-ity), year of birth fixed effects and place of birth fixed effects.

33

are purely observational in nature as we lack a strong research design for causal inference.38

Second, our calculations only provide reduced form estimates that may capture many other

factors that are associated with the channel in question.

Marriage Previous work has documented positive assortative mating on education in the

marriage market (Anukriti and Dasgupta, 2017; Behrman and Rosenzweig, 2002; Hahn et al.,

2018). We explore the impacts of INPRES on the spousal characteristics of the first gen-

eration men and women (Table 6) and find some evidence of improved marital outcomes

for women exposed to the program, but not for exposed men (Table 6, Panel B). Specifi-

cally, women fully exposed to the program are more likely to marry higher educated men.

However, the estimates are smaller and noisier in the expanded sample, which includes both

partially and fully exposed individuals as treated.

Our back of the envelope calculations suggest that the impact of INPRES on husband’s

education may explain up to 1.5% of the effect on women’s self-reported health (Table B.18).

In the case of marriage, we limit our investigation of the contribution of this channel to first

generation outcomes only. This is because we find that second generation outcomes are

causally influenced mainly by mother’s education and not father’s education, so it would not

make sense in the context of our particular analysis to consider marital sorting.39

Fertility A large literature has established that improving women’s education increases

the opportunity cost of having children. As a result, women may delay childbearing and

have fewer children due to the trade-off between child quantity and quality (Becker and

Lewis, 1973; Osili and Long, 2008).40 In addition, better educated women may make better

fertility choices through greater knowledge and more effective use of contraceptives (Kim,

38If there is omitted variable bias, then it is possible we may be overstating the effect of each mechanism.39For the second generation, our back of the envelope calculations suggest that the mother’s assortative

mating channel may explain up to 6% of the maternal exposure to INPRES impacts on children’s humancapital.

40See for example Aaronson et al. (2014); Hahn et al. (2018).

34

2010; Rosenzweig and Schultz, 1989).41

We investigate women’s fertility responses to INPRES exposure in Table 7. Using the

fertility history in the IFLS, we begin by examining whether women delayed their first

pregnancy and find a small negative effect that is statistically insignificant (Panel A, col. 1).

Next, we examine birth spacing between the first and second child and find no significant

effect (Panel A, col. 2). Finally, since our sample of first generation women have likely

completed their fertility by 2012-2014, column 3 presents the effects on total number of

children ever born. Although the estimated impact is negative, suggesting a decline in the

number of children, it is noisy, which may be due to the small sample of the IFLS. To assess

this hypothesis, we replicate our analysis in a larger nationally representative sample using

the 2014 Susenas which contains information on the number of births. Panel B and Panel C

present the results from this analysis for all the provinces in Indonesia and for the provinces

surveyed in the IFLS and IFLS-E respectively. We find that exposure to one INPRES school

per 1,000 children reduced the number of live births by 0.05 in all provinces (Panel B) and a

similar, but noisily estimated effect on fertility when we restrict the sample to the IFLS and

IFLS-E provinces (Panel C). The point estimates from the Susenas data are qualitatively

similar to those from the IFLS data, and more precise and statistically significant due to

the larger sample size. Breierova and Duflo (2004) examine the impacts of INPRES on

female fertility and find qualitatively similar effects to ours, but they use data from 1995

when treated cohorts may have not completed their fertility. The magnitude of the INPRES

effects on total fertility are similar to findings from previous studies of primary school reforms

41We examine the impact of INPRES exposure on women’s contraception knowledge and ever use ofcontraception and find no statistically significant effect. Ideally, we would have information on women’shistory of contraceptive use to analyze contraceptive use before the first birth.

35

in Nigeria (Osili and Long, 2008) and Uganda (Keats, 2018).42

One concern is whether the INPRES effects on intergenerational outcomes may be ex-

plained by fertility responses and the quantity-quality trade-off. Our back of the envelope

calculations indicate that the reduction in total fertility might explain between 2% and 8%

of the second-generation impacts on heath outcomes and test scores (Table B.18). This

suggests that, while we observe that women exposed to INPRES have fewer children, which

could then alter the second generation sample, this concern does not appear to be driving

our second-generation findings.43

Household resources Another channel we consider is whether the greater access to edu-

cation afforded by INPRES allows individuals to accumulate more resources and make more

productive investments in health (Grossman, 1972).44 To measure household resources, we

use data on per capita consumption, which is a widely used proxy for household income and

well-being.45 Specifically, we use the IFLS-5 (in 2014) and IFLS-E (in 2012) to measure log

per capita expenditure (in 2012 Rupiah), which is based on weekly or monthly per capita

42For example, exploiting the introduction of universal primary schooling in Nigeria, Osili and Long (2008)find that increasing female education by one year decreases early fertility by 0.26 births. Similarly, usingthe elimination of primary school fees in Uganga, Keats (2018) documented that an extra year of schoolingdecreases early fertility by 0.36 births. Considering the INPRES effect on years of schooling documentedby Breierova and Duflo (2004), our reduce-form estimate on number of births is similar in magnitude tothe findings from these studies. Breierova and Duflo (2004) find that exposure to INPRES increase yearsof schooling by 0.2 years for women in the restricted sample. We replicate similar but noisy effects in theappendix of our companion paper Mazumder et al. (2019). Putting together these INPRES impacts onyears of schooling and Osili and Long (2008)’s estimate of the effect of one year of female schooling on totalbirths would imply a fertility effect of INPRES of -0.05 births. Similarly, Keats (2018) estimated effectswould imply a fertility effect of -0.07 births. Our estimated effect of INPRES (-0.05) is similar to thosecalculations.

43Even if we consider a larger association between total fertility and children outcomes (estimate + 2SDs),fertility responses might explain between 3% and 10% of the intergenerational effects.

44We, unfortunately, are not able to analyze the role of productive investments in health due to the low rateof preventive care. We are also not able to examine the role of allocative efficiency (Kenkel, 1991) throughimproved knowledge due to data limitation. The IFLS includes questions on breast cancer awareness forwomen in the first generation, but the response rate is low.

45See for example Akee et al. (2018) who find that household income and resources are an important inputfor individual health and children’s human capital. Per capita consumption is recorded more precisely thanhousehold income and it is considered a better proxy for permanent income than current income (Groshet al., 2000).

36

food and non-food expenditure.46 In addition, we use data from the same survey years to

construct a housing quality index which we use as a proxy for the household environment.

The index combines poor housing characteristics, which include: poor toilets, floors, roofs,

walls and an indicator for high occupancy per room.47 We standardize each item by sub-

tracting the mean and dividing by the standard deviation of the comparison group, and

create an index that is the average of the standardized outcomes. Thus, higher values of the

index reflect poorer housing quality. We also considered an index of asset ownership which

includes: savings, vehicle, land, TV, appliances, refrigerator, and house.

We find that individuals who were exposed to INPRES do in fact have greater household

resources as captured by these measures (Table 8). Specifically, we show that each additional

INPRES school per 1,000 children leads to approximately 5% higher consumption (col. 1),

a lower index of poor housing quality of about 0.03 to 0.04 SDs (col. 2), and imprecisely

estimated increase in the asset index by about 0.02 SDs (col. 3).48 We generally find stronger

effects for women who were fully exposed to the program (Table 9, cols. 1-6).

We find that these improvements account for a small portion of our effects. Household

resources explain between 2% and 3% of the effect of INPRES on the long-term health index.

For the second generation outcomes, this channel can explain between 6% and 22% of the

intergenerational impacts on test scores and health outcomes (Table B.18).

Migration We also consider migration as a mechanism. INPRES may have led individuals

to move to either better or worse areas. For example, in the context of schools built for rural

blacks in the American South in the early 20th century, Aaronson et al. (2020) find that

a failure to separately account for the effects on migration to areas where blacks suffered

46We exclude annual non-food expenditure, which includes items such as land and vehicle purchases. TableB.12 presents the components separately.

47Poor toilet is captured by not having access to a toilet (including shared or public toilet). Poor floorincludes board or lumber, bamboo, or dirt floor. Poor roof includes leaves or wood. Poor wall includeslumber or board and bamboo or mat. High occupancy per room is defined as more than two persons perroom in the house (based on household size).

48We explore the role of each item of the index and find that the results are consistent across the housingitems (Tables B.13-B.14).

37

higher mortality can obscure the health promoting effects of education. To address this, we

directly consider how INPRES may have influenced migration decisions. In other words, we

examine whether individuals exposed to INPRES are more or less likely to migrate out of

their place of birth than those not exposed to the program. We explore several indicators of

migration below.

First generation: We analyze migration in the first generation using two indicators

and estimate equation 1 (Table 10, Panel A). First, we create an indicator that takes the

value one if the district of birth is different from the current district of residence in any of

the waves of the IFLS. Second, we create a similarly coded variable that only compares the

district of birth to the district of residence in 2012 (IFLS-E) or 2014 (IFLS-5), which are the

waves we use to measure our health outcomes of interest for the first generation.49 We find

that exposure to INPRES has no significant or sizable effect on either of these indicators

of migration. We also estimate equation 1 on the sample of non-movers and find similar

estimates on long-term health (Table B.15).

Second generation: We further explore migration by considering the possibility of

parents (first generation) migrating before or after the birth of their child (second generation)

by estimating equation 2 (Table 10, Panel B). For the former, we create an indicator that

takes the value one if the mother’s district of birth is different from the child’s district of birth.

For the latter, we create an indicator that takes the value one if the child’s district of birth is

different from the child’s current district of residence. We find that maternal exposure to the

INPRES program has no significant effect on migration before or after the birth of her child.

For robustness, we estimate the impacts of the second generation’s human capital outcomes

on the sample of non-movers and find similar intergenerational effects (Table B.15).

49The first measure is broader in that it accounts for some individuals who might have left their districtof birth but then returned, and individuals who were no longer in the sample in 2014. The second measureis directly comparable to our estimation sample. Due to the small sample size of the IFLS, we replicate ouranalysis using the 2014 Susenas and find qualitatively similar results (Table A.5).

38

Neighborhood quality Another potential channel through which schooling can impact

long-term and intergenerational outcomes is through human capital externalities. One

prominent example is that education could lead to higher levels of political participation

(Wantchekon et al., 2014), which could in turn lead to policies that improve socioeconomic

outcomes. Indeed, Martinez-Bravo (2017) found that INPRES increased the level of educa-

tion of potential candidates to local leadership positions, which then improved local gover-

nance and the provision of public goods. This finding suggests that we ought to examine

the role of neighborhood quality as a potential mechanism behind our results. There is also

strong evidence from the U.S. and Australia that children’s human capital is shaped by the

neighborhood where they grow up (Chetty and Hendren, 2018; Deutscher, 2019).

To investigate the potential role of neighborhoods, we use the IFLS community survey

(IFLS-5 in 2014 and IFLS-E in 2012) to create indices of education and health service pro-

vision at the community level (village or township in rural and urban areas respectively).

We also create a poverty index based on the fraction of households in the community in

the following social assistance programs: subsidized rice program (Raskin), subsidized na-

tional health insurance (Jamkesmas), subsidized regional health insurance (Jamkesda).50

The poverty index captures both poverty and access to anti-poverty programs.

Greater provision of educational services at the local level may explain the second gen-

eration’s improved educational outcomes, while access to health services may explain both

the first and second generation’s improved health outcomes. One caveat is that we are only

able to observe neighborhood quality for non-movers in IFLS-5 (in 2014 with respect to

IFLS-1 in 1993).51 Since migration does not appear to drive our estimated effects, selection

concerns are minimized. For each community in IFLS-5 (in 2014) and IFLS-E (in 2012), we

create an education index using the number of primary, junior high, and high schools used

50Raskin is a national program that provides rice, the staple food, at highly subsidized prices for poorhouseholds. Jamkesmas is also a national program that provides health coverage for poor households.Jamkesda is similar to Jamkesmas, but provided at the province or district level.

51The main IFLS collects information on the original 312 communities sampled in IFLS-1 in each wave ofthe survey. Therefore, neighborhood quality measures are only available for individuals who resided in theoriginal 312 communities in 2014.

39

by the community.52 Similarly, the health index includes: an indicator for having a majority

of the residents using piped water and a similar indicator for private toilet, the number of

community health centers, and the number of midwives available to the community.

We find no statistically significant effect on neighborhood access to education facilities

(Table 8, col. 4).53 In contrast, we find that individuals who were exposed to INPRES have

better access to health services in their communities, and this effect is statistically significant

for those fully exposed to INPRES in the restricted sample (Panel B, col. 5). In addition,

this effect seems to be concentrated among women (Table 9, cols. 7-12). Similarly, we find

that women in the restricted sample tend to reside in communities with lower poverty (by

0.069 standard deviations, col. 12). These results are in line with the INPRES program

improving public good provision.

When we examine the contribution of community health resources on the first generation’s

long-term health, our back of the envelope calculations suggest that the INPRES impact on

this channel may explain between 1% and 3% of the effect on self-reported health (Table

B.18). For the second generation, exposure to a neighborhood with better health facilities

can explain between 3% and 10% of the program effect on the second generation’s health

index and height for age.

Other mechanisms These results suggest that a significant portion of the first genera-

tion’s long-term health impacts and intergenerational human capital effects cannot be at-

tributed to the mechanisms that we are able to measure. We cautiously interpret these

findings as suggesting that the direct effects of INPRES on the educational attainment of

individuals in the first generation are likely to be the main explanation for the gains ob-

served in the second generation. Nevertheless, an important caveat for our analysis is that

there may be many other interesting and important potential mechanisms at play that we

52We standardize each item by subtracting the mean and dividing by the standard deviation of the com-parison group, and create a new summary index variable that is the simple average of all standardizedoutcomes.

53We estimate equations 1 and 2 and cluster the standard errors by district of birth (parental district ofbirth for the second generation) and community.

40

simply cannot measure. These include factors such as decision-making power, knowledge

acquisition, family human capital investments, and maternal mental health.54

8 Discussion and Conclusion

We find that increased access to primary education through a massive school construc-

tion program in Indonesia has important spillover effects both on other dimensions of human

capital, and on the offspring of individuals exposed to the program. Individuals exposed to

INPRES have better health outcomes 40 years later, including better self-reported health

status, fewer mental health symptoms, and fewer chronic conditions. These findings consti-

tute new evidence on the causal effects of education on health in the context of a developing

country.

We also find striking evidence of intergenerational spillovers. Children of mothers ex-

posed to INPRES are less likely to be stunted, have better self-reported health, and have

significantly higher test scores. We find no statistically significant effects through paternal

exposure and the point estimates are either similar or smaller than those from maternal

exposure. Our findings are not driven by migration responses and are robust to alternative

specifications. With respect to the role of fertility responses, while we find that women ex-

posed to the program experience a decline in their total fertility, our “back-of-the envelope”

calculations imply that this channel only explains up to 8 percent of our second generation

effects. Additionally, we present some evidence that greater household resources and bet-

ter neighborhood quality are potential mechanisms for our intergenerational findings. Our

findings point to one important way in which policy makers can influence human capital

outcomes even into the subsequent generation.

54There is growing causal evidence on the impacts of maternal mental health, especially during pregnancy,on children’s human capital and long-term outcomes (Black et al., 2016; Persson and Rossin-Slater, 2018).Additionally, in LMICs, maternal mental health has been identified as an important predictor of childdevelopment (Walker et al., 2011). We were unable to calculate associations between maternal mental healthand children’s human capital outcomes for the comparison group because the children of the comparisongroup are mainly observed in the earlier waves of the IFLS, where adult mental health was not measured.

41

How important are these spillovers and do they justify the kinds of large-scale expendi-

tures on school construction programs such as INPRES? We directly address this through a

cost benefit analysis. Specifically, we calculate the real internal rate of return of the program

both with and without taking into account the spillover effects. In order to make these cal-

culations, we make several simplifying assumptions. For example, we make the conservative

assumption that the INPRES schools were operational for twenty years and phased out by

1997.55 For benefits, we only include the earnings and health gains for each generation. We

use the results from existing studies in order to translate the magnitude of the health and

educational improvements we observe in the second generation on lifetime earnings. The

complete details are available in Appendix C.

We estimate that the internal rate of return of the INPRES program is about 7.9% when

only including the returns to primary school completion for the first generation. This is sim-

ilar both to what Duflo (2001) finds for INPRES as well as what Aaronson and Mazumder

(2011) find for the Rosenwald school construction program that took place in the rural Amer-

ican South early in the 20th century. When we include the long-term health improvements

in the first generation, the estimated internal rate of return rises to about 8.8%.56 When we

further account for the intergenerational benefits (health and test scores), we find a vastly

higher rate of return of as much as 24.8%. This suggests that traditional cost-benefit anal-

yses of this type of intervention that only take into account the returns to education may

significantly underestimate the societal benefits. These findings have highly salient policy

implications for countries that are still struggling with access to basic education and are

contemplating large-scale schooling interventions. Our results strongly suggest that these

nations should take into account the potential spillover gains to health and to subsequent

generations.

55Thus, cohorts born between 1963 and 1989 received benefits from the program.56Akresh et al. (2018) estimate the internal rate of return based on the first generation’s taxes and improved

living standards. They find a rate of return between 10.5% and 20.7%.

42

References

Aaronson, D., F. Lange, and B. Mazumder (2014). Fertility transitions along the extensive

and intensive margins. American Economic Review 104 (11), 3701–24.

Aaronson, D. and B. Mazumder (2011). The impact of rosenwald schools on black achieve-

ment. Journal of Political Economy 119 (5), 821–888.

Aaronson, D., B. Mazumder, S. G. Sanders, and E. J. Taylor (2020). Estimating the effect

of school quality on mortality in the presence of migration: Evidence from the jim crow

south. Journal of Labor Economics .

Aguero, J. M. and M. Ramachandran (2018). The intergenerational transmission of schooling

among the education-rationed. Journal of Human Resources , 0816–8143R.

Akee, R., W. Copeland, E. J. Costello, and E. Simeonova (2018). How does household

income affect child personality traits and behaviors? American Economic Review 108 (3),

775–827.

Akresh, R., D. Halim, and M. Kleemans (2018). Long-term and intergenerational effects

of education: Evidence from school construction in indonesia. Technical report, National

Bureau of Economic Research.

Almond, D., J. Currie, and V. Duque (2018). Childhood circumstances and adult outcomes:

Act ii. Journal of Economic Literature 56 (4), 1360–1446.

Anukriti, S. and S. Dasgupta (2017). Marriage markets in developing countries. In The

Oxford Handbook of Women and the Economy.

Asher, S., P. Novosad, and C. Rafkin (2020). Intergenerational mobility in india: New

methods and estimates across time, space, and communities. Working Paper .

Ashraf, N., N. Bau, N. Nunn, and A. Voena (2018). Bride price and female education.

Baird, S., E. Chirwa, J. de Hoop, and B. Ozler (2014). Girl power: Cash transfers and

adolescent welfare: Evidence from a cluster-randomized experiment in malawi. In African

Successes, Volume II: Human Capital, pp. 139–164. University of Chicago Press.

Baird, S., C. McIntosh, and B. Ozler (2011). Cash or condition? evidence from a cash

transfer experiment. The Quarterly journal of economics 126 (4), 1709–1753.

43

Barham, T. (2012). Enhancing cognitive functioning: Medium-term effects of a health

and family planning program in matlab. American Economic Journal: Applied Eco-

nomics 4 (1), 245–73.

Barham, T., K. Macours, and J. A. Maluccio (2013). More schooling and more learning?:

Effects of a three-year conditional cash transfer program in nicaragua after 10 years.

Technical report, Inter-American Development Bank.

Barr, A. and C. Gibbs (2018). Breaking the cycle? the intergenerational effects of head

start.

Becker, G. S. (1964). 1993. human capital. A Theoretical and Empirical Analysis with Special

References to Education.

Becker, G. S. and H. G. Lewis (1973). On the interaction between the quantity and quality

of children. Journal of political Economy 81 (2, Part 2), S279–S288.

Behrman, J. R. and M. R. Rosenzweig (2002). Does increasing women’s schooling raise the

schooling of the next generation? American Economic Review 92 (1), 323–334.

Bettinger, E., M. Kremer, M. Kugler, C. Medina, C. Posso, and J. E. Saavedra (2017). Can

educational voucher programs pay for themselves?

Bharati, T., S. Chin, and D. Jung (2016). Does education affect time preference?

Bharati, T., S. Chin, and D. Jung (2018). Recovery from an early life shock through improved

access to schools: Evidence from indonesia.

Black, S. E., P. J. Devereux, and K. G. Salvanes (2016). Does grief transfer across genera-

tions? bereavements during pregnancy and child outcomes. American Economic Journal:

Applied Economics 8 (1), 193–223.

Booth, A. (1998). The Indonesian economy in the nineteenth and twentieth centuries: A

history of missed opportunities. Springer.

Breierova, L. and E. Duflo (2004). The impact of education on fertility and child mortality:

Do fathers really matter less than mothers?

Chetty, R. and N. Hendren (2018). The impacts of neighborhoods on intergenerational

mobility ii: County-level estimates. The Quarterly Journal of Economics 133 (3), 1163–

1228.

44

Chou, S.-Y., J.-T. Liu, M. Grossman, and T. Joyce (2010). Parental education and child

health: evidence from a natural experiment in taiwan. American Economic Journal:

Applied Economics 2 (1), 33–61.

Clark, D. and H. Royer (2013). The effect of education on adult mortality and health:

Evidence from britain. American Economic Review 103 (6), 2087–2120.

Crespo, L., B. Lopez-Noval, and P. Mira (2014). Compulsory schooling, education, depression

and memory: New evidence from sharelife. Economics of Education Review 43, 36–46.

Currie, J. and E. Moretti (2003). Mother’s education and the intergenerational transmis-

sion of human capital: Evidence from college openings. The Quarterly journal of eco-

nomics 118 (4), 1495–1532.

Cutler, D. M., A. Lleras-Muney, and T. Vogl (2011). Socioeconomic status and health:

Dimensions and mechanisms. The Oxford Handbook of Health Economics .

De Chaisemartin, C. and X. d’Haultfoeuille (2018). Fuzzy differences-in-differences. The

Review of Economic Studies 85 (2), 999–1028.

DeSalvo, K. B., V. S. Fan, M. B. McDonell, and S. D. Fihn (2005). Predicting mortality and

healthcare utilization with a single question. Health services research 40 (4), 1234–1246.

Deutscher, N. (2019). Place, peers, and the teenage years: Long-run neighbourhood effects

in australia. forthcoming at American Economic Journal: Applied Economics .

Dinsa, G. D., Y. Goryakin, E. Fumagalli, and M. Suhrcke (2012). Obesity and socioeconomic

status in developing countries: a systematic review. Obesity reviews 13 (11), 1067–1079.

Duflo, E. (1999). Essays in empirical development economics. Ph. D. thesis, Massachusetts

Institute of Technology.

Duflo, E. (2001). Schooling and labor market consequences of school construction in in-

donesia: Evidence from an unusual policy experiment. American economic review 91 (4),

795–813.

Duflo, E. (2004). The medium run effects of educational expansion: Evidence from a large

school construction program in indonesia. Journal of Development Economics 74 (1), 163–

197.

Dursun, B., R. Cesur, and I. R. Kelly (2017). The value of mandating maternal education

in a developing country. Technical report, National Bureau of Economic Research.

45

Dursun, B., R. Cesur, and N. Mocan (2018). The impact of education on health outcomes and

behaviors in a middle-income, low-education country. Economics & Human Biology 31,

94–114.

East, C. N., S. Miller, M. Page, and L. R. Wherry (2017). Multi-generational impacts of

childhood access to the safety net: Early life exposure to medicaid and the next genera-

tion’s health.

Emran, S. and F. Shilpi (2015). Gender, geography and generations: Intergenerational

educational mobility in post-reform india. World Development 72, 362–380.

Emran, S. and F. Shilpi (2018). Estimating intergenerational mobility with incomplete data:

coresidency and truncation bias in rank-based relative and absolute mobility measures.

Policy Research Working Paper Series (8414).

Fernald, L. C., P. J. Gertler, and L. M. Neufeld (2008). Role of cash in conditional cash

transfer programmes for child health, growth, and development: an analysis of mexico’s

oportunidades. The Lancet 371 (9615), 828–837.

Frankenberg, E. and N. R. Jones (2004). Self-rated health and mortality: does the rela-

tionship extend to a low income setting? Journal of health and social behavior 45 (4),

441–452.

Friedman, W., M. Kremer, E. Miguel, and R. Thornton (2016). Education as liberation?

Economica 83 (329), 1–30.

Fuchs, V. R. (1980). Time preference and health: an exploratory study.

Galama, T., A. Lleras-Muney, and H. van Kippersluis (2018). The effect of education on

health and mortality: A review of experimental and quasi-experimental evidence. In

Oxford Research Encyclopedia of Economics and Finance.

Glewwe, P. and M. Kremer (2006). Schools, teachers, and education outcomes in developing

countries. Handbook of the Economics of Education 2, 945–1017.

Glewwe, P. and K. Muralidharan (2016). Improving education outcomes in developing coun-

tries: Evidence, knowledge gaps, and policy implications. In Handbook of the Economics

of Education, Volume 5, pp. 653–743. Elsevier.

Grepin, K. A. and P. Bharadwaj (2015). Maternal education and child mortality in zimbabwe.

Journal of health economics 44, 97–117.

46

Grosh, M., P. Glewwe, et al. (2000). Designing household survey questionnaires for devel-

oping countries. World Bank Publications .

Grossman, M. (1972). On the concept of health capital and the demand for health. Journal

of Political economy 80 (2), 223–255.

Hahn, Y., K. Nuzhat, and H.-S. Yang (2018). The effect of female education on marital

matches and child health in bangladesh. Journal of Population Economics 31 (3), 915–

936.

Heckman, J. J., J. E. Humphries, and G. Veramendi (2018). The nonmarket benefits of

education and ability. Journal of human capital 12 (2), 282–304.

Hoynes, H., D. W. Schanzenbach, and D. Almond (2016). Long-run impacts of childhood

access to the safety net. American Economic Review 106 (4), 903–34.

Idler, E. L. and Y. Benyamini (1997). Self-rated health and mortality: a review of twenty-

seven community studies. Journal of health and social behavior , 21–37.

Karachiwalla, N. and G. Palloni (2019). Human capital and structural transformation: Quasi-

experimental evidence from Indonesia, Volume 1836. Intl Food Policy Res Inst.

Kazianga, H., D. Levy, L. L. Linden, and M. Sloan (2013). The effects of” girl-friendly”

schools: Evidence from the bright school construction program in burkina faso. American

Economic Journal: Applied Economics 5 (3), 41–62.

Keats, A. (2018). Women’s schooling, fertility, and child health outcomes: Evidence from

uganda’s free primary education program. Journal of Development Economics 135, 142–

159.

Kenkel, D. S. (1991). Health behavior, health knowledge, and schooling. Journal of Political

Economy 99 (2), 287–305.

Kim, J. (2010). Women’s education and fertility: An analysis of the relationship between ed-

ucation and birth spacing in indonesia. Economic development and cultural change 58 (4),

739–774.

Kling, J. R., J. B. Liebman, and L. F. Katz (2007). Experimental analysis of neighborhood

effects. Econometrica 75 (1), 83–119.

Lindeboom, M., A. Llena-Nozal, and B. van Der Klaauw (2009). Parental education and child

health: Evidence from a schooling reform. Journal of health Economics 28 (1), 109–131.

47

Lundborg, P., A. Nilsson, and D.-O. Rooth (2014). Parental education and offspring out-

comes: evidence from the swedish compulsory school reform. American Economic Journal:

Applied Economics 6 (1), 253–78.

Martinez-Bravo, M. (2017). The local political economy effects of school construction in

indonesia. American Economic Journal: Applied Economics 9 (2), 256–89.

Mazumder, B. (2008). Does education improve health? a reexamination of the evidence

from compulsory schooling laws. Economic Perspectives 32 (2), 2–17.

Mazumder, B. (2012). The effects of education on health and mortality. Nordic Economic

Policy Review 1 (2012), 261–301.

Mazumder, B., M. Rosales-Rueda, and M. Triyana (2019). Intergenerational human capital

spillovers: Indonesia’s school construction and its effects on the next generation. Forth-

coming at AEA: Papers and Proceedings .

Nimubona, A.-D. and D. Vencatachellum (2007). Intergenerational education mobility of

black and white south africans. Journal of Population Economics 20 (1), 149–182.

Nores, M. and W. S. Barnett (2010). Benefits of early childhood interventions across the

world:(under) investing in the very young. Economics of education review 29 (2), 271–282.

Oreopoulos, P. (2007). Do dropouts drop out too soon? wealth, health and happiness from

compulsory schooling. Journal of public Economics 91 (11-12), 2213–2229.

Oreopoulos, P. and K. G. Salvanes (2011). Priceless: The nonpecuniary benefits of schooling.

Journal of Economic perspectives 25 (1), 159–84.

Osili, U. O. and B. T. Long (2008). Does female schooling reduce fertility? evidence from

nigeria. Journal of development Economics 87 (1), 57–75.

Parker, S. W. and T. Vogl (2018). Do conditional cash transfers improve economic outcomes

in the next generation? evidence from mexico. Technical report, National Bureau of

Economic Research.

Perez-Alvarez, M. (2017). Returns to cognitive skills in 7 developing countries.

Persson, P. and M. Rossin-Slater (2018). Family ruptures, stress, and the mental health of

the next generation. American economic review 108 (4-5), 1214–52.

Psacharopoulos, G. and H. A. Patrinos* (2004). Returns to investment in education: a

further update. Education economics 12 (2), 111–134.

48

Razzaque, A., A. Mustafa, and P. K. Streatfield (2014). Do self-reported health indicators

predict mortality? evidence from matlab, bangladesh. Journal of biosocial science 46 (5),

621–634.

Rosenzweig, M. R. and T. P. Schultz (1989). Schooling, information and nonmarket produc-

tivity: contraceptive use and its effectiveness. International Economic Review , 457–477.

Rossin-Slater, M. and M. Wust (2018). What is the added value of preschool? long-term

impacts and interactions with an infant health intervention.

Simes, R. J. (1986). An improved bonferroni procedure for multiple tests of significance.

Biometrika 73 (3), 751–754.

Sohn, K. (2015). The height premium in indonesia. Economics & Human Biology 16, 1–15.

Sohn, K. (2017). The fatter are happier in indonesia. Quality of Life Research 26 (2),

393–402.

Strauss, J., F. Witoelar, B. Sikoki, and A. Wattie (2016). User’s Guide for the Indonesia

Family Life Survey, Wave 5. RAND Santa Monica.

Triyana, M., F. Witoelar, J. Strauss, and P. Rukumnuaykit (2019). Smoking in a developing

country: the socio-economic determinants of smoking in indonesia. Technical report.

UNESCO (2015). Education for All 2000-2015: achievements and challenges; EFA global

monitoring report, 2015. UNESCO publishing.

Walker, S. P., T. D. Wachs, S. Grantham-McGregor, M. M. Black, C. A. Nelson, S. L.

Huffman, H. Baker-Henningham, S. M. Chang, J. D. Hamadani, B. Lozoff, et al. (2011).

Inequality in early childhood: risk and protective factors for early child development. The

lancet 378 (9799), 1325–1338.

Wantchekon, L., M. Klasnja, and N. Novta (2014). Education and human capital externali-

ties: evidence from colonial benin. The Quarterly Journal of Economics 130 (2), 703–757.

World Bank (1989). Indonesia: Basic education study.

Yamauchi, F. (2012). Prenatal seasonality, child growth, and schooling investments: Evi-

dence from rural indonesia. Journal of Development Studies 48 (9), 1323–1341.

Zha, D. (2019). Schooling expansion and marriage market: Evidence from indonesia.

49

Tables and Figures

First generation estimations

Figure 1: Effect on the first generation: poor health index

Notes: Overall poor health index corresponds to a summary index from the multiple self-reported healthmeasures analyzed: self-reported poor general health, number of days missed, any chronic conditions, numberof conditions, and mental health screening score where higher scores correspond to more symptoms ofdepression. The index has mean 0, SD 1 based on those born between 1950-1962 in low INPRES areas.Expanded sample includes those born between 1950-1972. Restricted sample includes those born between1957-1962 or 1968-1972. Covariates include: district, year of birth, month of birth, ethnicity (Javanesedummy), birth-year interacted with: the number of school-aged children in the district in 1971 (before thestart of the program), the enrollment rate of the district in 1971 and the exposure of the district to anotherINPRES program: a water and sanitation program. Robust standard errors clustered at district of birth.Bars indicate the 95% confidence intervals. Significance: *** p < 0.01, ** p < 0.05, * p < 0.10.

50

Figure 2: Effect on the first generation: poor health index event study

-.1-.0

50

.05

.1#

of In

pres

Sch

ools

X y

ob

Age

24-2

1

Age

20-1

8

Age

17-1

5

Age

14-1

2

Age

11-9

Age

8-6

Age

5-2

Mai

n D

ID m

odel

Age in 1974 when INPRES started

Notes: Overall poor health index corresponds to a summary index from the multiple self-reported healthmeasures analyzed: self-reported poor general health, number of days missed, any chronic conditions, numberof conditions, and mental health screening score where higher scores correspond to more symptoms ofdepression. The index has mean 0, SD 1 based on those born between 1950-1962 in low INPRES areas.Expanded sample includes those born between 1950-1972. Restricted sample includes those born between1957-1962 or 1968-1972. Covariates include: district, year of birth, month of birth, ethnicity (Javanesedummy), birth-year interacted with: the number of school-aged children in the district in 1971 (before thestart of the program), the enrollment rate of the district in 1971 and the exposure of the district to anotherINPRES program: a water and sanitation program. Robust standard errors clustered at district of birth.Bars indicate the 95% confidence intervals. Significance: *** p < 0.01, ** p < 0.05, * p < 0.10.

51

Figure 3: Intergenerational effects: poor health index

-0.03***

-0.02-.0

6-.0

4-.0

20

.02

SD

Mother exp Father exp.

Panel A. Expanded sample

Each parent separate

-0.03*-0.02

-.06

-.04

-.02

0.0

2SD

Mother exp. Father exp.

Panel B. Restricted sample

-0.04**

-0.00

-.06-

.04-

.02

0.0

2.0

4SD

Mother exp Father exp.

Panel C. Expanded sample

Both parents

-0.03

-0.01

-.06

-.04

-.02

0.0

2.0

4SD

Mother exp. Father exp.

Panel D. Restricted sample

Notes: Overall poor health index corresponds to a summary index from the following health measuresfor the second generation: being stunted, anemic and self-reported poor health. Sample corresponds tochildren born to first generation INPRES individuals. Expanded sample includes children born to adults bornbetween 1950-1972. Restricted sample includes children born to adults born between 1957-1962 or 1968-1972.Covariates include the following FE: parent year of birth×1971 enrollment, parent year of birth×1971 numberof children, parent year of birth×water sanitation program, child’s gender, birth order, year and month ofbirth dummies, urban, and ethnicity (Javanese dummy). Panel A and B robust standard errors in parenthesesclustered at each parent’s district of birth. Panel C and D robust standard errors in parentheses clusteredtwo-way at mother and father district of birth. Bars indicate the 95% confidence intervals. Significance: ***p < 0.01, ** p < 0.05, * p < 0.10.

52

Figure 4: Intergenerational outcome: effect on poor health index

Panel A. Maternal exposure

-.1-.0

50

.05

.1m

othe

r # o

f Inp

res

Scho

ols

X yo

b

Age

24-2

1

Age

20-1

8

Age

17-1

5

Age

14-1

2

Age

11-9

Age

8-6

Age

5-2

mot

her m

ain

DID

mod

el

mother year of birth

Poor health Index

-.2-.1

0.1

.2m

othe

r # o

f Inp

res

Scho

ols

X yo

b

Age

24-2

1

Age

20-1

8

Age

17-1

5

Age

14-1

2

Age

11-9

Age

8-6

Age

5-2

mot

her m

ain

DID

mod

el

mother year of birth

Secondary Exam

Event study second generation mother exposure

Panel B. Paternal exposure

-.1-.0

50

.05

.1fa

ther

# o

f Inp

res

Scho

ols

X yo

b

Age

24-2

1

Age

20-1

8

Age

17-1

5

Age

14-1

2

Age

11-9

Age

8-6

Age

5-2

fath

er m

ain

DID

mod

el

father year of birth

Poor health Index

-.3-.2

-.10

.1.2

fath

er #

of I

npre

s Sc

hool

s X

yob

Age

24-2

1

Age

20-1

8

Age

17-1

5

Age

14-1

2

Age

11-9

Age

8-6

Age

5-2

fath

er m

ain

DID

mod

el

father year of birth

Secondary Exam

Event study second generation father exposure

Notes: Overall poor health index corresponds to a summary index from the following health measuresfor the second generation: being stunted, anemic and self-reported poor health. Sample corresponds tochildren born to first generation INPRES individuals. Expanded sample includes children born to adultsborn between 1950-1972. Restricted sample includes children born to adults born between 1957-1962 or 1968-1972. Covariates include the following FE: parent year of birth×1971 enrollment, parent year of birth×1971number of children, parent year of birth×water sanitation program, child’s gender, birth order, year andmonth of birth dummies, urban, and ethnicity (Javanese dummy). Robust standard errors in parenthesestwo-way clustered at the parent’s district of birth and individual level. Bars indicate the 95% confidenceintervals. Significance: *** p < 0.01, ** p < 0.05, * p < 0.10.

53

Table 1: Program effect on the first generation’s health outcomes

Panel A. Expanded sample

(1) (2) (3) (4) (5) (6)

Poor health Self-reported Days Any chronic No. of Mental

index healthy missed condition conditions health

Born bet. 1963-1972 -0.041*** 0.039*** -0.171** -0.025** -0.047** -0.180

× INPRES (0.013) (0.010) (0.068) (0.011) (0.019) (0.141)

No. of obs. 9891 10792 10420 10729 10729 10244

Dep. var. mean -0.024 0.677 2.010 0.482 0.833 5.402

R-squared 0.10 0.08 0.06 0.09 0.10 0.09

Joint test p-value 0.0003

Panel B. Restricted sample

(1) (2) (3) (4) (5) (6)

Poor health Self-reported Days Any chronic No. of Mental

index healthy missed condition conditions health

Born bet. 1968-1972 -0.038** 0.033*** -0.079 -0.014 -0.037 -0.271*

× INPRES (0.016) (0.012) (0.102) (0.017) (0.024) (0.153)

No. of obs. 5537 5999 5826 5940 5940 5741

Dep. var. mean -0.058 0.705 1.887 0.456 0.772 5.378

R-squared 0.12 0.09 0.07 0.10 0.11 0.12

Joint test p-value 0.045

Notes: Overall poor health index corresponds to a summary index from the multiple self-reported healthmeasures analyzed: self-reported poor general health, number of days missed, any chronic conditions, numberof conditions, and mental health screening score where higher scores correspond to more symptoms ofdepression. The index has mean 0, SD 1 based on those born between 1950-1962 in low INPRES areas.Healthy takes the value one if a respondent reports being “Very healthy” or “Healthy”. “Days missed”corresponds to the number of days a respondent missed his or her activities due to health reasons in the past4 weeks prior to the survey. Any chronic condition and the number of chronic conditions come from self-reported chronic conditions. Mental health score is based on the following items: being bothered by things,having trouble concentrating, feeling depressed, feeling like everything was an effort, feeling hopeful aboutthe future, feeling fearful, having restless sleep, feeling happy, lonely, and unable to get going. Higher scorescorrespond to poorer mental health. Expanded sample includes those born between 1950-1972. Restrictedsample includes those born between 1957-1962 or 1968-1972. Covariates include: district, year of birth,month of birth, ethnicity (Javanese dummy), birth-year interacted with: the number of school-aged childrenin the district in 1971 (before the start of the program), the enrollment rate of the district in 1971 and theexposure of the district to another INPRES program: a water and sanitation program. Robust standarderrors clustered at district of birth. Significance: *** p < 0.01, ** p < 0.05, * p < 0.10.

54

Tab

le2:

Pro

gram

effec

ton

the

firs

tge

ner

atio

n’s

hea

lth

outc

omes

by

gender

Pan

elA

.E

xp

an

ded

sam

ple

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

(10)

(11)

(12)

Poor

hea

lth

Sel

f-re

port

edD

ays

mis

sed

Any

chro

nic

No.

of

Men

tal

ind

exh

ealt

hy

con

dit

ion

con

dit

ion

sh

ealt

h

Mal

eF

emal

eM

ale

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Bor

nb

et.

1963

-197

2-0

.021

-0.0

58**

*0.

017

0.0

62***

-0.1

16

-0.2

05*

-0.0

15

-0.0

39***

-0.0

30

-0.0

75**

-0.1

09

-0.2

35

×IN

PR

ES

(0.0

21)

(0.0

19)

(0.0

13)

(0.0

14)

(0.0

86)

(0.1

12)

(0.0

20)

(0.0

14)

(0.0

26)

(0.0

34)

(0.1

88)

(0.2

38)

No.

ofob

s.48

1750

7452

965496

5140

5280

5266

5463

5266

5463

4962

5282

Dep

.va

r.m

ean

-0.1

250.

069

0.70

90.6

47

1.7

17

2.2

84

0.4

02

0.5

56

0.6

80

0.9

76

5.1

29

5.6

53

R-s

qu

ared

0.12

0.11

0.13

0.1

00.0

90.0

80.1

00.1

00.1

20.1

10.1

20.1

2

Gen

der

diff

eren

ce0.

176

0.0

08

0.5

17

0.3

65

0.2

67

0.6

70

p-v

alu

e

Pan

elB

.R

estr

icte

dsa

mp

le

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

(10)

(11)

(12)

Poor

hea

lth

Sel

f-re

port

edD

ays

mis

sed

Any

chro

nic

No.

of

Men

tal

ind

exh

ealt

hy

con

dit

ion

con

dit

ion

sh

ealt

h

Mal

eF

emal

eM

ale

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Bor

nb

et.

1968

-197

2-0

.010

-0.0

61**

0.00

60.0

60***

-0.0

25

-0.1

02

0.0

03

-0.0

32*

-0.0

09

-0.0

72*

0.0

15

-0.5

20**

×IN

PR

ES

(0.0

32)

(0.0

26)

(0.0

19)

(0.0

17)

(0.1

25)

(0.1

79)

(0.0

31)

(0.0

18)

(0.0

40)

(0.0

40)

(0.2

81)

(0.2

14)

No.

ofob

s.26

6828

6929

303069

2861

2965

2900

3040

2900

3040

2751

2990

Dep

.va

r.m

ean

-0.1

710.

041

0.74

20.6

73

1.5

96

2.1

51

0.3

61

0.5

41

0.5

78

0.9

45

5.0

58

5.6

55

R-s

qu

ared

0.15

0.14

0.14

0.1

20.1

20.1

20.1

30.1

30.1

40.1

40.1

60.1

5

Gen

der

diff

eren

ce0.

089

0.0

20

0.7

33

0.3

00

0.2

87

0.1

37

p-v

alu

e

Note

s:O

ver

all

poor

hea

lth

ind

exco

rres

pon

ds

toa

sum

mary

ind

exfr

om

the

mu

ltip

lese

lf-r

eport

edh

ealt

hm

easu

res

an

aly

zed

:se

lf-r

eport

edp

oor

gen

eral

hea

lth

,nu

mb

erof

days

mis

sed

,any

chro

nic

con

dit

ion

s,nu

mb

erof

con

dit

ion

s,an

dm

enta

lh

ealt

hsc

reen

ing

score

wh

ere

hig

her

score

sco

rres

pon

dto

more

sym

pto

ms

of

dep

ress

ion

.T

he

ind

exh

as

mea

n0,

SD

1b

ase

don

those

born

bet

wee

n1950-1

962

inlo

wIN

PR

ES

are

as.

Th

ep

-valu

eof

the

join

tte

stfo

rit

ems

inth

eh

ealt

hin

dex

for

men

inP

an

elA

is0.5

45,

an

din

Pan

elB

is0.9

92.

Th

ep

-valu

eof

the

join

tte

stfo

rit

ems

inth

eh

ealt

hin

dex

for

wom

enin

Pan

elA

is<

0.0

01,

an

din

Pan

elB

is0.0

02.

Hea

lthy

takes

the

valu

eon

eif

are

spon

den

tre

port

sb

ein

g“V

ery

hea

lthy”

or

“H

ealt

hy”.

’Days

mis

sed

’co

rres

pon

ds

toth

enu

mb

erof

days

are

spon

den

tm

isse

dh

isor

her

act

ivit

ies

du

eto

hea

lth

reaso

ns

inth

ep

ast

4w

eeks

pri

or

toth

esu

rvey

.A

ny

chro

nic

con

dit

ion

an

dth

enu

mb

erof

chro

nic

con

dit

ion

sco

me

from

self

-rep

ort

edch

ron

icco

nd

itio

ns.

Men

tal

hea

lth

score

isb

ase

don

the

follow

ing

item

s:b

ein

gb

oth

ered

by

thin

gs,

havin

gtr

ou

ble

con

centr

ati

ng,

feel

ing

dep

ress

ed,

feel

ing

like

ever

yth

ing

was

an

effort

,fe

elin

gh

op

efu

lab

ou

tth

efu

ture

,fe

elin

gfe

arf

ul,

havin

gre

stle

sssl

eep

,fe

elin

gh

ap

py,

lon

ely,

an

du

nab

leto

get

goin

g.

Hig

her

score

sco

rres

pon

dto

poore

rm

enta

lh

ealt

h.

Exp

an

ded

sam

ple

incl

ud

esth

ose

born

bet

wee

n1950-1

972.

Res

tric

ted

sam

ple

incl

ud

esth

ose

born

bet

wee

n1957-1

962

or

1968-1

972.

Covari

ate

sin

clu

de:

dis

tric

t,yea

rof

bir

th,

month

of

bir

th,

eth

nic

ity

(Javan

ese

du

mm

y),

bir

th-y

ear

inte

ract

edw

ith

:th

enu

mb

erof

sch

ool-

aged

child

ren

inth

ed

istr

ict

in1971

(bef

ore

the

start

of

the

pro

gra

m),

the

enro

llm

ent

rate

of

the

dis

tric

tin

1971

an

dth

eex

posu

reof

the

dis

tric

tto

an

oth

erIN

PR

ES

pro

gra

m:

aw

ate

ran

dsa

nit

ati

on

pro

gra

m.

Rob

ust

stan

dard

erro

rscl

ust

ered

at

dis

tric

tof

bir

th.

Sig

nifi

can

ce:

***

p<

0.0

1,

**

p<

0.0

5,

*p<

0.1

0.

55

Second generation estimations

Table 3: Program effects on intergenerational outcomes: poor health index

Panel A. Expanded sample

(1) (2) (3)

Mother Father Both

only only parents

Mother bet. 1963-72 -0.034*** -0.037**

×INPRES (0.011) (0.015)

Father bet. 1963-72 -0.017 -0.003

×INPRES (0.012) (0.017)

No. of obs. 17919 17384 19042

R-squared 0.17 0.15 0.17

Panel B. Restricted sample

(1) (2) (3)

Mother Father Both

only only parents

Mother bet. 1968-72 × INPRES -0.027* -0.026

(0.016) (0.019)

Father bet. 1968-72 × INPRES -0.019 -0.010

(0.016) (0.019)

No. of obs. 9911 9007 13707

R-squared 0.17 0.15 0.17

Notes: Overall poor health index corresponds to a summary index from the following health measures for thesecond generation: being stunted, anemic and self-reported poor health. Sample corresponds to children bornto first generation INPRES individuals. Expanded sample includes children born to adults born between1950-1972. Restricted sample includes children born to adults born between 1957-1962 or 1968-1972. Forcolumn 1 and 2, covariates include the following FE: parent year of birth×1971 enrollment, parent year ofbirth×1971 number of children, parent year of birth×water sanitation program, child’s gender, birth order,year and month of birth dummies, urban, and ethnicity (Javanese dummy). Robust standard errors inparentheses clustered at the parent’s district of birth. For column 3 (both parents), the sample correspondsto children either born mothers or fathers in the first generation sample. These models include mother’sand father’s exposure and the full set of covariates for the mother, while for the father we include: provinceof birth, two-year bins for year of birth fixed effects and interactions between the father’s year of birth (intwo-year bins) and the father’s district-level covariates. In this estimation, standard errors are clusteredtwo-way at the mother and father district of birth. Significance: *** p < 0.01, ** p < 0.05, * p < 0.10

56

Tab

le4:

Pro

gram

effec

tson

inte

rgen

erat

ional

outc

omes

:H

ealt

hm

easu

res

Pan

elA

.E

xp

an

ded

sam

ple

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

(10)

(11)

(12)

Hei

ght-

for-

age

Stu

nti

ng

Sel

f-re

port

edh

ealt

hy

An

emia

Mot

her

Fat

her

Bot

hM

oth

erF

ath

erB

oth

Moth

erF

ath

erB

oth

Moth

erF

ath

erB

oth

only

only

par

ents

on

lyon

lyp

are

nts

on

lyon

lyp

are

nts

on

lyon

lyp

are

nts

Mot

her

bet

.19

63-7

20.

056*

0.06

5*

-0.0

25*

-0.0

23

0.0

12**

0.0

13*

-0.0

10

-0.0

13

×IN

PR

ES

(0.0

30)

(0.0

33)

(0.0

13)

(0.0

16)

(0.0

05)

(0.0

07)

(0.0

11)

(0.0

13)

Fat

her

bet

.19

63-7

2-0

.016

-0.0

44

-0.0

09

0.0

02

0.0

04

-0.0

01

-0.0

09

-0.0

07

×IN

PR

ES

(0.0

29)

(0.0

40)

(0.0

12)

(0.0

16)

(0.0

05)

(0.0

07)

(0.0

10)

(0.0

11)

No.

ofob

s.21

382

1989

022

224

21382

19890

22224

18648

18171

19852

18104

17535

19217

Dep

.va

r.m

ean

-1.6

48-1

.627

-1.6

62

0.3

65

0.3

60

0.3

70

0.8

33

0.8

70

0.8

31

0.2

56

0.2

47

0.2

50

R-s

qu

ared

0.16

0.17

0.18

0.1

20.1

30.1

30.3

10.2

50.2

90.0

70.0

70.0

7

Mot

her

join

tte

stp

-val

0.01

Fat

her

join

tte

stp

-val

0.16

Pan

elB

.R

estr

icte

dsa

mp

le

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

(10)

(11)

(12)

Hei

ght-

for-

age

Stu

nti

ng

Sel

f-re

port

edh

ealt

hy

An

emia

Mot

her

Fat

her

Bot

hM

oth

erF

ath

erB

oth

Moth

erF

ath

erB

oth

Moth

erF

ath

erB

oth

only

only

par

ents

on

lyon

lyp

are

nts

on

lyon

lyp

are

nts

on

lyon

lyp

are

nts

Mot

her

bet

.19

68-7

20.

044

0.04

6-0

.022

-0.0

08

0.0

05

0.0

07

-0.0

11

-0.0

19

×IN

PR

ES

(0.0

46)

(0.0

46)

(0.0

20)

(0.0

21)

(0.0

06)

(0.0

08)

(0.0

14)

(0.0

16)

Fat

her

bet

.19

68-7

2-0

.055

-0.0

35

-0.0

07

-0.0

13

-0.0

05

-0.0

04

-0.0

18

-0.0

09

×IN

PR

ES

(0.0

49)

(0.0

49)

(0.0

17)

(0.0

19)

(0.0

06)

(0.0

08)

(0.0

12)

(0.0

12)

No.

ofob

s.11

464

1000

715

562

11464

10007

15562

10319

9423

14303

9994

9084

13821

Dep

.va

r.m

ean

-1.6

44-1

.641

-1.6

51

0.3

68

0.3

65

0.3

70

0.8

53

0.8

98

0.8

52

0.2

50

0.2

44

0.2

42

R-s

qu

ared

0.19

0.20

0.20

0.1

40.1

50.1

40.2

90.1

90.2

60.0

90.0

90.0

9

Mot

her

join

tte

stp

-val

0.45

Fat

her

join

tte

stp

-val

0.14

Note

s:S

am

ple

corr

esp

on

ds

toch

ild

ren

born

tofi

rst

gen

erati

on

INP

RE

Sin

div

idu

als

.E

xp

an

ded

sam

ple

incl

ud

esch

ild

ren

born

toad

ult

sb

orn

bet

wee

n1950-1

972.

Res

tric

ted

sam

ple

incl

ud

esch

ild

ren

born

toad

ult

sb

orn

bet

wee

n1957-1

962

or

1968-1

972.

For

the

each

pare

nt

mod

els

covari

ate

sin

clu

de

the

foll

ow

ing

FE

:p

are

nt

yea

rof

bir

th×

1971

enro

llm

ent,

pare

nt

yea

rof

bir

th×

1971

nu

mb

erof

child

ren

,p

are

nt

yea

rof

bir

th×

wate

rsa

nit

ati

on

pro

gra

m,

chil

d’s

gen

der

,b

irth

ord

er,

yea

ran

dm

onth

of

bir

thd

um

mie

s,u

rban,

an

det

hn

icit

y(J

avan

ese

du

mm

y).

Rob

ust

stan

dard

erro

rsin

pare

nth

eses

clu

ster

edat

the

pare

nt’

sd

istr

ict

of

bir

th.

For

the

both

pare

nts

mod

els,

the

sam

ple

corr

esp

on

ds

toch

ild

ren

eith

erb

orn

moth

ers

or

fath

ers

inth

efi

rst

gen

erati

on

sam

ple

.T

hes

em

od

els

incl

ud

em

oth

er’s

an

dfa

ther

’sex

posu

rean

dth

efu

llse

tof

covari

ate

sfo

rth

em

oth

er,

wh

ile

for

the

fath

erw

ein

clu

de:

pro

vin

ceof

bir

th,

two-y

ear

bin

sfo

ryea

rof

bir

thfi

xed

effec

tsan

din

tera

ctio

ns

bet

wee

nth

efa

ther

’syea

rof

bir

th(i

ntw

o-y

ear

bin

s)an

dth

efa

ther

’sd

istr

ict-

level

covari

ate

s.In

this

esti

mati

on

,st

an

dard

erro

rsare

clu

ster

edtw

o-w

ay

at

the

moth

eran

dfa

ther

dis

tric

tof

bir

th.

Sig

nifi

can

ce:

***

p<

0.0

1,

**

p<

0.0

5,

*p<

0.1

0.

57

Table 5: Program effects on intergenerational outcomes: National secondary test scores

Panel A. Expanded sample

(1) (2) (3)

Mother Father Both

only only parents

Mother bet. 1963-72 × INPRES 0.083*** 0.110***

(0.031) (0.039)

Father bet. 1963-72 ×INPRES 0.047 0.006

(0.033) (0.043)

No. of obs. 6819 5744 6604

R-squared 0.11 0.11 0.14

Panel B. Restricted sample

(1) (2) (3)

Mother Father Both

only only parents

Mother bet. 1968-72 × INPRES 0.105* 0.112**

(0.059) (0.057)

Father bet. 1968-72 × INPRES 0.103 0.040

(0.065) (0.056)

No. of obs. 3512 2639 4361

R-squared 0.14 0.17 0.17

Notes: Ninth grade test scores standardized for each exam year. Sample corresponds to children born to firstgeneration INPRES individuals. Expanded sample includes children born to adults born between 1950-1972.Restricted sample includes children born to adults born between 1957-1962 or 1968-1972. For column 1and 2, covariates include the following FE: parent year of birth×1971 enrollment, parent year of birth×1971number of children, parent year of birth×water sanitation program, child’s gender, birth order, year andmonth of birth dummies, urban, and ethnicity (Javanese dummy). Robust standard errors in parenthesesclustered at the parent’s district of birth. For column 3 (both parents), the sample corresponds to childreneither born mothers or fathers in the first generation sample. These models include mother’s and father’sexposure and the full set of covariates for the mother, while for the father we include: birth province, yearof birth, interactions between year of birth (in two-year bins) and the district-level covariates: the numberof school-aged children in the district in 1971 (before the start of the program), the enrollment rate of thedistrict in 1971 and the exposure of the district to the contemporaneous water and sanitation program. Inthis estimation, standard errors are clustered two-way at the mother and father district of birth. Significance:*** p < 0.01, ** p < 0.05, * p < 0.10.

58

Potential mechanisms

Table 6: Potential mechanisms: First generation’s marriage outcomes

Panel A. Expanded sample

(1) (2) (3) (4) (5) (6)

Spouse with primary Spouse with secondary Age difference

completion completion Respondent−spouse

Men Women Men Women Men Women

Born bet. 1963-72 0.007 0.006 -0.025 0.018 0.079 0.197

× INPRES (0.013) (0.013) (0.016) (0.016) (0.153) (0.271)

No. of obs. 5989 5884 5989 5884 5999 5443

Dep. var. mean 0.80 0.78 0.47 0.48 -5.14 4.93

R-squared 0.22 0.19 0.26 0.24 0.09 0.11

Panel B. Restricted sample

(1) (2) (3) (4) (5) (6)

Spouse with primary Spouse with secondary Age difference

completion completion Respondent−spouse

Men Women Men Women Men Women

Born bet. 1968-72 0.008 0.032** -0.012 0.052** -0.191 0.326

× INPRES (0.019) (0.016) (0.021) (0.024) (0.218) (0.362)

No. of obs. 3281 3226 3281 3226 3290 3004

Dep. var. mean 0.83 0.81 0.51 0.52 -4.89 4.80

R-squared 0.25 0.21 0.27 0.25 0.11 0.14

Notes: Primary completion corresponds to 6 years of education, secondary completion corresponds to 9years of education, and age difference is in years. Expanded sample includes those born between 1950-1972.Restricted sample includes those born between 1957-1962 or 1968-1972. Covariates include: district, year ofbirth, month of birth, ethnicity (Javanese dummy), birth-year interacted with: the number of school-agedchildren in the district in 1971 (before the start of the program), the enrollment rate of the district in 1971and the exposure of the district to another INPRES program: a water and sanitation program. Robuststandard errors clustered at district of birth. Significance: *** p < 0.01, ** p < 0.05, * p < 0.10.

59

Table 7: Potential mechanisms: Fertility

Panel A. IFLS and IFLS-E data

(1) (2) (3) (4) (5) (6)

Age 1st preg. Spacing 1st and 2nd child No. of births

Expanded Restricted Expanded Restricted Expanded Restricted

sample sample sample sample sample sample

Young cohort -0.182 -0.111 0.355 0.381 -0.072 -0.089

× INPRES (0.192) (0.264) (1.517) (1.963) (0.083) (0.080)

No. of obs. 5673 3110 4693 2553 5987 3293

Dep. var. mean 22.76 22.06 48.52 47.63 3.77 3.34

R-squared 0.17 0.21 0.15 0.18 0.33 0.32

Panel B. Susenas: All provinces

No. of births

Expanded Restricted

sample sample

Young cohort -0.048** -0.052*

× INPRES (0.023) (0.027)

No. of obs. 128,853 69,900

Dep. var. mean 4.20 3.95

R-squared 0.16 0.16

Panel C. Susenas: Restricted to IFLS and IFLS-E provinces

No. of births

Expanded Restricted

sample sample

Young cohort -0.037 -0.030

× INPRES (0.028) (0.033)

No. of obs. 99,727 53,739

Dep. var. mean 4.14 3.91

R-squared 0.17 0.17

Notes: Expanded sample includes those born between 1950-1975. Restricted sample includes those bornbetween 1957-1962 or 1968-1975. Young cohort for the expanded sample corresponds to those born between1963 and 1972. Young cohort for the restricted sample corresponds to those born between 1968 and 1972.Covariates include: district, year of birth, month of birth, ethnicity (Javanese dummy), birth-year interactedwith: the number of school-aged children in the district in 1971 (before the start of the program), theenrollment rate of the district in 1971 and the exposure of the district to another INPRES program: a waterand sanitation program. Robust standard errors clustered at district of birth. Significance: *** p < 0.01,** p < 0.05, * p < 0.10. 60

Table 8: Potential mechanisms: Household resources and neighborhood quality

Panel A. Expanded sample

(1) (2) (3) (4) (5) (6)

Household resources Neighborhood quality

Per capita PoorAsset index

Education Health Poverty

expenditure housing index index index

Born bet. 1963-1972 0.046** -0.031* 0.024 0.014 0.024 -0.003

× INPRES (0.018) (0.018) (0.028) (0.021) (0.022) (0.022)

No. of obs. 11941 11507 11951 9329 9078 8015

Dep. var. mean 12.860 -0.038

R-squared 0.15 0.38 0.20 0.56 0.48 0.50

Panel B. Restricted sample

(1) (2) (3) (4) (5) (6)

Household resources Neighborhood quality

Per capita PoorAsset index

Education Health Poverty

expenditure housing index index index

Born bet. 1968-1972 0.054** -0.040* 0.037 -0.001 0.055** -0.012

× INPRES (0.026) (0.020) (0.029) (0.031) (0.027) (0.027)

No. of obs. 6752 6502 6756 5118 4974 4377

Dep. var. mean 12.899 -0.042

R-squared 0.15 0.38 0.19 0.58 0.48 0.51

Notes: Total log per capita expenditure in 2012-14 based on weekly or monthly per capita food and non-food expenditure in 2012Rupiah). We exclude annual non-food expenditure (which includes items like land/vehicle purchases). Index of poor housingquality in 2012-14 includes: poor toilet, poor floor, poor roof, poor wall, high occupancy per room. Poor toilet is capturedby not having access to a toilet (including shared or public toilet). Poor floor includes board/lumber, bamboo, or dirt floor.Poor roof includes leaves or wood. Poor wall includes lumber/board and bamboo/mat. Occupancy per room is defined as morethan two persons per room in the house (based on household size). Asset index includes the following asset ownership: savings,vehicle, land, TV, appliances, refrigerator, and house. The community index is only available for households that continue toreside in the original IFLS enumeration areas. Education index includes the number of primary, junior high, and high schoolsused by the community. Health index includes the following: an indicator for having a majority of the residents using pipedwater, an indicator for having a majority of the residents using private toilet, the number of community health centers, andthe number of midwives available to the community. Poverty index includes the fraction of households in the community in thesubsidized rice program (Raskin), subsidized national health insurance (Jamkesmas), subsidized regional (province or district)health insurance (Jamkesda). Expanded sample includes those born between 1950-1972. Restricted sample includes those bornbetween 1957-1962 or 1968-1972. Covariates include: district, year of birth, month of birth, ethnicity (Javanese dummy),birth-year interacted with: the number of school-aged children in the district in 1971 (before the start of the program), theenrollment rate of the district in 1971 and the exposure of the district to another INPRES program: a water and sanitationprogram. Robust standard errors clustered at district of birth. Significance: *** p < 0.01, ** p < 0.05, * p < 0.10.

61

Tab

le9:

Pot

enti

alm

echan

ism

s:H

ouse

hol

dre

sourc

esan

dnei

ghb

orhood

qual

ity

by

gender

Pan

elA

.E

xp

an

ded

sam

ple

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

(10)

(11)

(12)

Per

cap

ita

Poor

hou

sin

gA

sset

ind

exE

du

cati

on

ind

exH

ealt

hin

dex

Pov

erty

ind

ex

exp

end

itu

re

Mal

eF

emal

eM

ale

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Bor

nb

et.

1963

-197

20.

044*

0.05

1-0

.036

-0.0

31

0.0

24

0.0

28

0.0

59

-0.0

03

0.0

36

0.0

14

0.0

15

-0.0

22

×IN

PR

ES

(0.0

23)

(0.0

34)

(0.0

25)

(0.0

22)

(0.0

34)

(0.0

35)

(0.0

38)

(0.0

29)

(0.0

30)

(0.0

32)

(0.0

35)

(0.0

33)

No.

ofob

s.60

0759

345785

5722

6007

5944

4526

4776

4399

4651

3920

4067

Dep

.va

r.m

ean

12.9

1212

.806

-0.0

45

-0.0

30

R-s

qu

ared

0.15

0.19

0.3

80.4

20.5

00.4

90.5

70.5

80.5

00.4

90.5

10.5

3

Pan

elB

.R

estr

icte

dsa

mp

le

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

(10)

(11)

(12)

Per

cap

ita

Poor

hou

sin

gA

sset

ind

exE

du

cati

on

ind

exH

ealt

hin

dex

Pov

erty

ind

ex

exp

end

itu

re

Mal

eF

emal

eM

ale

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Bor

nb

et.

1968

-197

20.

058*

0.06

0-0

.028

-0.0

70***

0.0

24

0.0

74**

0.0

22

0.0

08

0.0

36

0.0

79*

0.0

17

-0.0

69*

×IN

PR

ES

(0.0

33)

(0.0

42)

(0.0

36)

(0.0

23)

(0.0

43)

(0.0

35)

(0.0

42)

(0.0

46)

(0.0

35)

(0.0

44)

(0.0

36)

(0.0

40)

No.

ofob

s.34

0833

443276

3226

3407

3349

2464

2617

2389

2549

2122

2218

Dep

.va

r.m

ean

12.9

3812

.858

-0.0

40

-0.0

44

R-s

qu

ared

0.15

0.21

0.4

00.4

00.2

10.2

40.5

90.6

00.5

10.5

00.5

30.5

3

Not

es:

Tot

allo

gp

erca

pit

aex

pen

dit

ure

in20

12-1

4b

ase

don

wee

kly

or

month

lyp

erca

pit

afo

od

an

dn

on

-food

exp

end

itu

rein

2012Rupiah

).W

eex

clu

de

annu

aln

on-f

ood

exp

end

itu

re(w

hic

hin

clu

des

item

sli

kela

nd

/ve

hic

lep

urc

hase

s).

Ind

exof

poor

hou

sin

gqu

ali

tyin

2012-1

4in

clu

des

:p

oor

toil

et,

poor

floor

,p

oor

roof

,p

oor

wal

l,h

igh

occ

up

ancy

per

room

.P

oor

toil

etis

cap

ture

dby

not

hav

ing

acc

ess

toa

toil

et(i

ncl

ud

ing

share

dor

pu

bli

cto

ilet

).P

oor

floor

incl

ud

esb

oard

/lu

mb

er,

bam

boo,

ord

irt

floor.

Poor

roof

incl

ud

esle

aves

or

wood

.P

oor

wall

incl

ud

eslu

mb

er/b

oard

an

db

am

boo/m

at.

Occ

up

an

cyp

erro

omis

defi

ned

asm

ore

than

two

per

son

sp

erro

om

inth

eh

ou

se(b

ase

don

hou

seh

old

size

).A

sset

ind

exin

clu

des

the

foll

owin

gass

etow

ner

ship

:sa

vin

gs,

veh

icle

,la

nd

,T

V,

app

lian

ces,

refr

iger

ator,

an

dh

ou

se.

Th

eco

mm

un

ity

ind

exis

on

lyav

ail

able

for

hou

seh

old

sth

at

conti

nu

eto

resi

de

inth

eor

igin

alIF

LS

enu

mer

atio

nar

eas.

Ed

uca

tion

ind

exin

clu

des

the

nu

mb

erof

pri

mary

,ju

nio

rh

igh

,an

dh

igh

sch

ools

use

dby

the

com

mu

nit

y.H

ealt

hin

dex

incl

ud

esth

efo

llow

ing:

anin

dic

ator

for

hav

ing

am

ajo

rity

of

the

resi

den

tsu

sin

gp

iped

wate

r,an

ind

icato

rfo

rh

avin

ga

ma

jori

tyof

the

resi

den

tsu

sin

gp

riva

teto

ilet

,th

enu

mb

erof

com

mu

nit

yh

ealt

hce

nte

rs,

an

dth

enu

mb

erof

mid

wiv

esav

ail

ab

leto

the

com

mu

nit

y.P

over

tyin

dex

incl

ud

esth

efr

act

ion

ofh

ouse

hol

ds

inth

eco

mm

un

ity

inth

esu

bsi

diz

edri

cep

rogra

m(R

askin

),su

bsi

diz

edn

ati

on

al

hea

lth

insu

ran

ce(Jamkesm

as)

,su

bsi

diz

edre

gio

nal

(pro

vin

ceor

dis

tric

t)h

ealt

hin

sura

nce

(Jamkesda

).E

xp

an

ded

sam

ple

incl

ud

esth

ose

born

bet

wee

n1950-1

972.

Res

tric

ted

sam

ple

incl

ud

esth

ose

born

bet

wee

n19

57-1

962

or19

68-1

972.

Cov

aria

tes

incl

ud

e:d

istr

ict,

yea

rof

bir

th,

month

of

bir

th,

eth

nic

ity

(Jav

an

ese

du

mm

y),

bir

th-y

ear

inte

ract

edw

ith

:th

enu

mb

erof

sch

ool

-age

dch

ild

ren

inth

ed

istr

ict

in1971

(bef

ore

the

start

of

the

pro

gra

m),

the

enro

llm

ent

rate

of

the

dis

tric

tin

1971

an

dth

eex

posu

reof

the

dis

tric

tto

anot

her

INP

RE

Sp

rogr

am:

aw

ate

ran

dsa

nit

ati

on

pro

gra

m.

Rob

ust

stan

dard

erro

rscl

ust

ered

at

dis

tric

tof

bir

th.

Sig

nifi

can

ce:

***

p<

0.01

,**

p<

0.05

,*

p<

0.10

.

62

Table 10: Potential mechanisms: Migration

Panel A. First generation(1) (2) (3) (4)

Expanded sample Restricted SampleEver moved Moved by 2012-14 Ever moved Moved by 2012-14

Young cohort -0.002 0.005 -0.000 0.001× INPRES (0.010) (0.010) (0.012) (0.012)No. of obs. 14049 13199 7818 7356Dep. var. mean 0.342 0.296 0.345 0.297R-squared 0.25 0.25 0.25 0.25

Panel B. Second generation(1) (2) (3) (4)

Expanded sample Restricted SampleMaternal mig. pre Maternal mig. post Maternal mig. pre Maternal mig. post

Mother: Young cohort -0.006 0.014 0.022 0.033× INPRES (0.016) (0.016) (0.022) (0.027)No. of obs. 15464 15397 8117 8083Dep. var. mean 0.285 0.285 0.280 0.271R-squared 0.28 0.11 0.30 0.14

Notes: Ever moved is an indicator that takes the value one if the adult respondent’s district of birth isdifferent from the respondent’s current district of residence. Moved by 2012-14 is indicator that comparesthe respondent’s district of birth and his or her district of residence in 2012 (IFLS-E) or 2014 (IFLS).Maternal mig. pre is indicator that takes the value one if the mother’s district of birth is different from thechild’s district of birth. Maternal mig. post is an indicator that takes the value one if the child’s district ofbirth is different from the child’s current district of birth. Expanded sample includes those born between1950-1975. Restricted sample includes those born between 1957-1962 or 1968-1975. Young cohort for theexpanded sample corresponds to those born between 1963 and 1972. Young cohort for the restricted samplecorresponds to those born between 1968 and 1972. See Table 1 for covariates. Robust standard errorsclustered at district of birth. Significance: *** p < 0.01, ** p < 0.05, * p < 0.10.

63

Appendix

A Data Appendix

A.1 Coverage of the IFLS and INPRES program

Program intensity

We compare the intensity of the INPRES school construction project in the IFLS and

IFLS-E against the national record. The IFLS provinces include 13 out of Indonesia’s 26

provinces in 1993. They include: North Sumatra, West Sumatra, South Sumatra,

Lampung, Jakarta, West Java, Central Java, Yogyakarta, East Java, Bali, West Nusa

Tenggara, South Kalimantan, and South Sulawesi. The IFLS-E provinces include the

following 7 provinces in 2012: East Nusa Tenggara, East Kalimantan, Southeast Sulawesi,

Maluku, North Maluku, West Papua, and Papua. The IFLS and IFLS-E include almost

300 of Indonesia’s 519 districts. Figure A.1 shows the intensity of the INRES program at

the national level while Figure A.2 shows the intensity of the INPRES program in the IFLS

and IFLS-E districts. A comparison of Figures A.1 and A.2 shows that the IFLS and

IFLS-E include both high and low intensity program districts.

Figure A.1: INPRES exposure - All Indonesia

Sources: Esri, HERE, Garmin, Intermap, increment P Corp., GEBCO, USGS, FAO, NPS, NRCAN, GeoBase, IGN, Kadaster NL, Ordnance Survey, Esri Japan, METI, Esri China (Hong Kong), swisstopo, © OpenStreetMap contributors,and the GIS User Community

INPRES Exposure

0-1.5

1.5-2.5

2.5-4.0

4.0-6.5

6.5-9.0

Source: Authors’ calculations based on Duflo (2001)

64

Figure A.2: INPRES exposure in the IFLS and IFLS-E districts

Sources: Esri, HERE, Garmin, Intermap, increment P Corp., GEBCO, USGS, FAO, NPS, NRCAN, GeoBase, IGN, Kadaster NL, Ordnance Survey, Esri Japan, METI, Esri China (Hong Kong), swisstopo, © OpenStreetMap contributors,and the GIS User Community

INPRES Exposure

0-1.5

1.5-2.5

2.5-4.0

4.0-6.5

6.5-9.0

Source: Authors’ calculations based on the IFLS, IFLS-E, and Duflo (2001)

A.2 Data construction

Indonesia is administratively divided into provinces, districts (regencies or cities),

sub-districts, and villages in rural areas or townships in urban areas. The IFLS over

sampled urban and rural areas outside of the main island of Java. IFLS-1 included 7,224

households residing in 13 of Indonesia’s 26 provinces in 1993. These households resided in

approximately 200 districts, which corresponded to 321 enumeration areas in 312

communities. A community is defined as a village in rural areas and a township in urban

areas. The IFLS-E includes 2,500 households residing in seven provinces in eastern

Indonesia, which corresponded to about 50 districts and 99 communities. Households in

the main IFLS and IFLS-E resided in almost 300 of Indonesia’s 514 districts.

Date and district of birth

To obtain the sample of first generation individuals, we begin by identifying individuals

who were born between 1950 and 1972 in the IFLS and IFLS-E. In each wave, the IFLS

household roster includes information on date of birth (month and year). Also, the IFLS

asks respondents over the age of 15 their place of birth in the wave in which they first join

the survey. Indonesia experienced district proliferation over time, so we match each district

to the 1993 district code in IFLS1. INPRES school construction in the district, water and

65

sanitation program, enrollment in 1971, number of school-aged children in 1971: We obtain

these variables from Duflo (2001).

Linking the first and second generation

To identify the second generation, who are the children of the first generation individuals,

we use the household relationship in the household roster and women’s birth history,

matched to the household roster. In each wave, the survey includes an individual’s

relationship to the head of the household, and an identifier for an individual’s mother and

father if the mother and father are in the same household. The IFLS also includes a

woman’s birth history, which allows us to match mothers to their children, and

subsequently to children’s outcomes.

Long-term outcomes for the first generation

Good self-reported health takes the value one if a respondent reported his or her self status

as “very healthy” or “healthy”. The literature in epidemiology has established that

self-reported health status is a valid and comprehensive health measure that is highly

predictive of well-known health markers such as mortality in both high and lower income

countries, even after controlling for socio-demographic factors (DeSalvo et al., 2005; Idler

and Benyamini, 1997; Razzaque et al., 2014). As additional adult health outcomes, we

include the number of days a respondent missed his or her activities in the past 4 weeks

prior to the survey. Respondents were also asked to report diagnosed chronic conditions,

and we use an indicator for any condition as well as the number of conditions. These

conditions include: hypertension, diabetes, tuberculosis, asthma, other respiratory

conditions, stroke, heart disease, liver condition, cancer, arthritis, high cholesterol,

depression/psychiatric condition.

To assess mental health, respondents were administered a series of 10 questions on how

frequently they experienced symptoms of depression using the CES-D. The items include

being bothered by things, having trouble concentrating, feeling depressed, feeling like

everything was an effort, feeling hopeful about the future, feeling fearful, having restless

sleep, feeling happy, lonely, and unable to get going. Each item includes 4 possible

responses: rarely or none in the past week, 1-2 days, 3-4 days, 5-7 days. The intensity of

each negative symptom is scored from 0 (rarely or none) to 3 (5-7 days a week). We recode

feeling hopeful about the future and feeling happy to reflect the negative symptoms. We

use the sum of the scores based on reported symptoms, where higher scores indicate a

higher likelihood of having depression.

66

Intergenerational outcomes

Using children’s height and age, we calculate height-for-age z-score using the WHO

reference data.57 Stunting takes the value one if a child’s height-for-age is more than two

standard deviations below the mean. Using children’s hemoglobin count, sex, and age, we

identify children with anemia. Specifically, anemia is defined as having a count of less than

11.5 grams of hemoglobin per deciliter (gr/dL) for children under 12 years of age. For

children between the ages of 12 and 15, the threshold is 12 gr/dL. The threshold is 12

gr/dL for girls over the age of 14 and 13 gr/dL for boys over the age of 14. Self-reported

health takes the value one if the child is reported as being healthy or very healthy.

INPRES exposure variable

For the first generation, the IFLS asks respondents over the age of 15 their place of birth in

the wave in which they first joined the survey. Additionally, in 2000, IFLS-3 asked all

respondents over the age of 15 their district of birth. We combine both sources of

information to obtain the respondents’ district of birth.

For the second generation, we identify mother-child and father-child pairs based on the

relationships within the household. We use mother-child (father-child) pairs by including

respondents identified as the biological child of adult female (male) respondents who were

born between 1950 and 1972. Additionally, in cases where the child’s place and/or date of

birth is missing from the household roster, we use women’s pregnancy history to identify

children born to women who were born between 1950 and 1972.

57We use the 2007 WHO growth chart, which is applicable to children between 0 and 19 years of age.

67

Table A.1: Summary statistics

(1) (2) (3)

Mean SD N

Panel A. First generation

Male 0.502 0.500 12,137

Born between 1963-1972 0.568 0.495 12,158

INPRES schools per 1,000 2.138 1.251 12,158

Javanese 0.455 0.498 12,158

Year of birth 1,963 6.392 12,158

Primary completion 0.674 0.469 12,158

Self-reported healthy 0.719 0.450 10,801

Days missed activities 1.841 3.267 10,429

Any chronic condition 0.422 0.494 10,738

No. of chronic conditions 0.683 1.007 10,738

Mental health screening score 5.513 4.694 10,254

Panel B. Second generation

First child 0.406 0.491 10,396

Male child 0.495 0.500 10,337

Child’s year of birth 1,988 6.554 10,396

Javanese 0.445 0.497 10,402

Mother born 1963-1972 0.445 0.497 10,396

Father born 1963-1972 0.484 0.500 10,396

Height for age z-score -1.649 1.096 25,482

Stunted 0.366 0.482 25,482

Anemia 0.249 0.433 21,952

Self-reported health 0.879 0.326 22,748

Notes: Summary statistics for the expanded sample, which includes first generation individuals born between1950-1972 and their children. The summary statistics for the health outcomes correspond to individualsobserved in the Wave 5 of the IFLS and IFLS-E. Second generation height captures multiple observationsper child as the IFLS measures height in all waves.

68

A.3 Data Comparison

Comparison of non-coresident and coresident second generation individuals in

the IFLS

The longitudinal nature of the main IFLS allows us to track non-coresident children who

make up the second generation. Non-coresident children in the IFLS account for about

40% of the total sample of the second generation individuals in our sample. Including only

coresident children in the second generation individuals may introduce some sample

selection. To address this, we explore the characteristics of coresident and non-coresident

children (Table A.2). We begin by comparing the raw sample mean, followed by the

adjusted differences. The adjusted difference takes into account mother’s district of birth

(in columns 9-10), and then we include mother’s year of birth (in columns 11-12). These

comparisons suggest that coresident children are more likely to be younger, come from

households with a higher asset index, and their parents have higher education.

Additionally, co-resident children are more likely to have mothers who were older at the

time of their first birth.

Comparison of coresident second generation individuals in the IFLS and a

nationally representative survey

We also examine the representativeness of the coresident children’s characteristics in the

IFLS by comparing the IFLS against the 2014 Socioeconomic survey, 2014 Susenas (Table

A.3). Comparing the characteristics of the national sample to the sample restricted to the

IFLS and IFLS-E provinces (columns 4-6 of Table A.3) suggests that the child

characteristics are similar to children’s characteristics in the national sample. Additionally,

when we restrict the 2014 Susenas to the IFLS provinces (columns 7-9 of Table A.3), the

children’s characteristics are similar to the co-resident children in the IFLS in columns 4-6

of Table A.2.

69

Tab

leA

.2:

Com

par

ison

ofnon

-cor

esid

ent

and

core

siden

tch

ildre

nin

the

IFL

S

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

(10)

(11)

(12)

Mot

her

’sbir

thM

other

’sbir

thpla

ce,

Out

ofH

ouse

hol

dIn

Hou

sehol

dR

awpla

ceF

Eye

arof

bir

thF

EM

ean

SD

Obs

Mea

nSD

Obs

Diff

SE

Mea

nD

iffSE

Mea

nD

iffSE

Child

mal

e0.

490.

502,

959

0.56

0.50

4,22

70.

064*

**0.

010.

06**

*0.

010.

07**

*0.

01C

hild

age

in20

1430

.08

5.77

2,61

224

.70

6.25

4,24

7-5

.449

***

0.17

-5.4

5***

0.17

-3.5

9***

0.15

Log

HH

exp

endit

ure

14.7

20.

752,

669

14.8

10.

744,

171

0.11

0***

0.02

0.07

***

0.02

0.08

***

0.02

Ass

etin

dex

0.11

0.95

2,66

90.

220.

764,

184

0.11

7***

0.03

0.08

***

0.03

0.10

***

0.03

Hou

sehol

dsi

ze3.

531.

872,

695

5.93

2.43

4,24

72.

443*

**0.

082.

33**

*0.

082.

52**

*0.

08C

hild:

pri

mar

yco

mple

te0.

730.

442,

956

0.69

0.46

4,19

0-0

.047

***

0.01

-0.0

4***

0.01

-0.0

4***

0.01

Child:

seco

ndar

yco

mple

te0.

530.

502,

944

0.53

0.50

4,18

1-0

.014

0.01

-0.0

00.

01-0

.01

0.01

Mot

her

:pri

mar

yco

mple

te0.

750.

432,

977

0.81

0.39

4,24

70.

069*

**0.

010.

06**

*0.

010.

03**

*0.

01F

ather

:pri

mar

yco

mple

te0.

830.

382,

782

0.85

0.36

3,98

60.

034*

**0.

010.

03**

0.01

0.02

0.01

Mot

her

:ye

ars

ofsc

hool

ing

7.10

3.83

2,97

77.

943.

954,

247

0.96

3***

0.12

0.82

***

0.11

0.47

***

0.10

Fat

her

:ye

ars

ofsc

hool

ing

8.38

4.05

2,78

28.

934.

053,

986

0.73

9***

0.13

0.62

***

0.11

0.41

***

0.12

Urb

an0.

700.

462,

695

0.65

0.48

4,24

7-0

.010

0.02

-0.0

5***

0.02

-0.0

5***

0.02

Mot

her

bor

n19

63-7

20.

330.

472,

977

0.53

0.50

4,24

70.

213*

**0.

010.

21**

*0.

01-0

.00

0.00

Mot

her

’sye

arof

bir

th1,

959.

635.

672,

977

1,96

2.60

5.89

4,24

73.

123*

**0.

183.

04**

*0.

18-0

.00

0.00

Fat

her

’sye

arof

bir

th1,

953.

927.

842,

782

1,95

7.56

8.01

3,99

33.

698*

**0.

263.

57**

*0.

260.

43**

*0.

16E

ver

mov

ed0.

120.

322,

689

0.03

0.17

4,24

5-0

.082

***

0.01

-0.0

8***

0.01

-0.0

9***

0.01

Mot

her

’sag

eat

firs

tbir

th21

.70

4.23

1,25

023

.48

4.51

1,50

41.

857*

**0.

171.

73**

*0.

182.

29**

*0.

18

Note

s:O

bse

rvati

on

sin

cols

.1-6

wei

ghte

dby

IFL

Scr

oss

sect

ion

al

wei

ght.

Sta

nd

ard

erro

rscl

ust

ered

at

moth

er’s

dis

tric

tof

bir

th.

Sig

nifi

can

ce:

***

p<

0.0

1,

**

p<

0.0

5,

*p

<0.1

0.

70

Tab

leA

.3:

Cor

esid

ent

childre

nch

arac

teri

stic

sin

Suse

nas

2014

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

All

pro

vin

ces

IFL

San

dIF

LS-E

pro

vin

ces

IFL

Spro

vin

ces

Mea

nSD

Obs

Mea

nSD

Obs

Mea

nSD

Obs

Child

mal

e0.

560.

5017

7,43

60.

550.

5013

5,69

80.

550.

5010

5,94

8C

hild

pri

mar

yco

mple

te0.

860.

3417

4,13

00.

870.

3413

3,07

50.

870.

3310

4,48

5C

hild

seco

ndar

yco

mple

te0.

640.

4817

4,13

00.

650.

4813

3,07

50.

650.

4810

4,48

5C

hild

year

ofbir

th1,

993.

487.

3417

7,43

61,

993.

447.

3713

5,69

81,

993.

347.

3710

5,94

8H

ouse

hol

dsi

ze5.

231.

8517

7,43

65.

211.

8413

5,69

85.

151.

8010

5,94

8M

other

bor

n19

63-7

20.

650.

4817

7,43

60.

650.

4813

5,69

80.

650.

4810

5,94

8M

other

’sag

eat

mar

riag

e19

.76

4.01

177,

433

19.7

24.

0313

5,69

519

.62

4.01

105,

945

Mot

her

’sye

arof

bir

th1,

964.

225.

8317

7,43

61,

964.

205.

8313

5,69

81,

964.

185.

8410

5,94

8M

other

:pri

mar

yco

mple

te0.

750.

4316

0,55

50.

750.

4312

2,49

50.

750.

4397

,105

Mot

her

:se

condar

yco

mple

te0.

380.

4916

0,55

50.

380.

4912

2,49

50.

380.

4997

,105

Mot

her

:ye

ars

ofsc

hool

ing

7.49

3.79

160,

288

7.48

3.79

122,

296

7.47

3.80

96,9

50U

rban

0.54

0.50

177,

436

0.57

0.50

135,

698

0.59

0.49

105,

948

Not

es:

Ob

serv

atio

ns

wei

ghte

dby

Su

sen

ascr

oss-

sect

ion

alw

eigh

t.

71

Comparison of primary school effects with a nationally representative survey

The IFLS contains rich information, but its relatively small sample may be a concern. To

address this, we examine the representativeness of the IFLS and IFLS-E by comparing our

data against the 2014 Socioeconomic survey, 2014 Susenas, since the fifth wave of the IFLS

was administered in 2014. The Susenas is a nationally representative survey that is

administered annually. The survey covers every district in Indonesia and includes some

questions that are also available in the IFLS and IFLS-E. Table A.4 presents primary

completion rates using the 2014 Susenas. Panel A presents estimated effects for all

provinces, Panel B presents effects for the IFLS and IFLS-E provinces, and Panel C

presents effects for the IFLS provinces. The effects are similar in panels A and B, and these

estimated effects are similar to our main estimates on first generation primary school

completion shown in Table B.1.

72

Table A.4: Primary completion: 2014 Susenas

Panel A. Susenas 2014: all provinces

(1) (2) (3) (4) (5) (6)

Expanded sample Restricted sample

All Male Female All Male Female

Young cohort 0.013*** 0.016*** 0.011** 0.017*** 0.017*** 0.017***

× INPRES (0.004) (0.004) (0.004) (0.004) (0.004) (0.005)

No. of obs. 236,270 122,213 114,057 130,855 67,674 63,181

Dep. var. mean (unweighted) 0.77 0.81 0.73 0.77 0.81 0.73

Dep. var. mean (weighted) 0.76 0.80 0.72 0.76 0.80 0.72

R-squared 0.099 0.084 0.107 0.099 0.088 0.111

Panel B. Susenas 2014: restricted to the IFLS and IFLS-E provinces

(1) (2) (3) (4) (5) (6)

Expanded sample Restricted sample

All Male Female All Male Female

Young cohort 0.017*** 0.019*** 0.015** 0.022*** 0.022*** 0.025***

× INPRES (0.005) (0.005) (0.006) (0.005) (0.005) (0.007)

No. of obs. 180,283 92,552 87,731 99,308 50,992 48,316

Dep. var. mean (unweighted) 0.77 0.81 0.73 0.77 0.81 0.73

Dep. var. mean (weighted) 0.76 0.76 0.72 0.76 0.76 0.72

R-squared 0.103 0.086 0.111 0.104 0.088 0.117

Notes: Expanded sample includes those born between 1950-1972. Restricted sample includes those bornbetween 1957-1962 or 1968-1972. Young cohort for the expanded sample corresponds to those bornbetween 1963 and 1972. Young cohort for the restricted sample corresponds to those born between 1968and 1972. Covariates include: district, year of birth, month of birth, ethnicity (Javanese dummy),birth-year interacted with: the number of school-aged children in the district in 1971 (before the start ofthe program), the enrollment rate of the district in 1971 and the exposure of the district to anotherINPRES program: a water and sanitation program. Robust standard errors clustered at district of birth.Significance: *** p < 0.01, ** p < 0.05, * p < 0.10.

We also compare migration between the IFLS and IFLS-E and 2014 Susenas in Table A.5.

Migration is defined as currently residing in a district that is different from one’s district of

73

birth. The estimated effects are similar in Panels A and B, and the estimates are similar to

our main estimates.

Table A.5: Potential mechanisms: Migration

(1) (2)

Current residence different from birth place

Expanded sample Restricted sample

Panel A. IFLS and IFLS-E data

Young cohort 0.005 0.001

× INPRES (0.010) (0.012)

No. of obs. 13,199 7,356

Dep. var. mean 0.296 0.297

R-squared 0.25 0.25

Panel B. Susenas: All provinces

Young cohort 0.000 0.001

× INPRES (0.002) (0.003)

No. of obs. 258,308 141,048

Dep. var. mean 0.250 0.254

R-squared 0.115 0.117

Panel C. Susenas: Restricted to IFLS and IFLS-E provinces

Young cohort 0.004 0.001

× INPRES (0.003) (0.003)

No. of obs. 198,337 141,048

Dep. var. mean 0.229 0.234

R-squared 0.125 0.117

Notes: Expanded sample includes those born between 1950-1975. Restricted sample includes those bornbetween 1957-1962 or 1968-1975. Young cohort for the expanded sample corresponds to those bornbetween 1963 and 1972. Young cohort for the restricted sample corresponds to those born between 1968and 1972. See Table 1 for covariates. Ever migrate takes the value one if respondent is not currentlyresiding in his/her district of birth. Robust standard errors clustered at district of birth. Significance: ***p < 0.01, ** p < 0.05, * p < 0.10.

74

B Additional figures and tables

Figure B.1: Pre-trends raw data: Primary school completion - IFLS

Notes: Primary completion rates for cohorts born between 1935 and 1958 from the main IFLS and IFLS-E.

Figure B.2: Pre-trends raw regression: Primary school completion - IFLS

Notes: Coefficients from difference-in-differences model that interacts the number of INPRES schools and

year of birth for cohorts born between 1945 and 1962. year of birth 1945 is the omitted category. Bars

indicate the 95% confidence intervals.

75

Figure B.3: Pre-trends raw regression: First generation’s health outcomes

Panel A. Males

-.2-.1

0.1

.2.3

Num

ber o

f Inp

res

Scho

ols

X ye

ar o

f birt

h

1945

1946

1947

1948

1949

1950

1951

1952

1953

1954

1955

1956

1957

1958

1959

1960

1961

1962

year of birth

Self-reported healthy

-10

12

3N

umbe

r of I

npre

s Sc

hool

s X

year

of b

irth

1945

1946

1947

1948

1949

1950

1951

1952

1953

1954

1955

1956

1957

1958

1959

1960

1961

1962

year of birth

Days missed activities

-.4-.2

0.2

.4N

umbe

r of I

npre

s Sc

hool

s X

year

of b

irth

1945

1946

1947

1948

1949

1950

1951

1952

1953

1954

1955

1956

1957

1958

1959

1960

1961

1962

year of birth

Any chronic condition-1

-.50

.51

Num

ber o

f Inp

res

Scho

ols

X ye

ar o

f birt

h

1945

1946

1947

1948

1949

1950

1951

1952

1953

1954

1955

1956

1957

1958

1959

1960

1961

1962

year of birth

Number of chronic cond.-3

-2-1

01

2N

umbe

r of I

npre

s Sc

hool

s X

year

of b

irth

1945

1946

1947

1948

1949

1950

1951

1952

1953

1954

1955

1956

1957

1958

1959

1960

1961

1962

year of birth

Mental health score

-.4-.2

0.2

.4N

umbe

r of I

npre

s Sc

hool

s X

year

of b

irth

1945

1946

1947

1948

1949

1950

1951

1952

1953

1954

1955

1956

1957

1958

1959

1960

1961

1962

year of birth

Poor health index

Health outcomes - Males

Panel B. Females

-.4-.2

0.2

.4N

umbe

r of I

npre

s Sc

hool

s X

year

of b

irth

1945

1946

1947

1948

1949

1950

1951

1952

1953

1954

1955

1956

1957

1958

1959

1960

1961

1962

year of birth

Self-reported healthy

-4-2

02

Num

ber o

f Inp

res

Scho

ols

X ye

ar o

f birt

h

1945

1946

1947

1948

1949

1950

1951

1952

1953

1954

1955

1956

1957

1958

1959

1960

1961

1962

year of birth

Days missed activities

-.2-.1

0.1

.2.3

Num

ber o

f Inp

res

Scho

ols

X ye

ar o

f birt

h

1945

1946

1947

1948

1949

1950

1951

1952

1953

1954

1955

1956

1957

1958

1959

1960

1961

1962

year of birth

Any chronic condition

-.50

.51

Num

ber o

f Inp

res

Scho

ols

X ye

ar o

f birt

h

1945

1946

1947

1948

1949

1950

1951

1952

1953

1954

1955

1956

1957

1958

1959

1960

1961

1962

year of birth

Number of chronic cond.

-4-2

02

4N

umbe

r of I

npre

s Sc

hool

s X

year

of b

irth

1945

1946

1947

1948

1949

1950

1951

1952

1953

1954

1955

1956

1957

1958

1959

1960

1961

1962

year of birth

Mental health score

-.4-.2

0.2

.4N

umbe

r of I

npre

s Sc

hool

s X

year

of b

irth

1945

1946

1947

1948

1949

1950

1951

1952

1953

1954

1955

1956

1957

1958

1959

1960

1961

1962

year of birth

Poor health index

Health outcomes - Females

Notes: Coefficients from difference-in-differences model that interacts the number of INPRES schools andyear of birth for cohorts born between 1945 and 1962. Bars indicate the 95% confidence intervals.

76

Figure B.4: Placebo regression: First generation

Notes: Sample includes individuals born between 1950 and 1962. “Placebo exposed group” for individualsborn between 1957 and 1962. Coefficients reported in standard deviation units. Overall health index corre-sponds to a summary index from the multiple self-reported health measures analyzed: self-reported generalhealth, days missed, if chronic conditions, number of conditions and mental health screening score. Healthindex has mean 0, SD 1 based on those born between 1950-1962 in low INPRES areas. Expanded sampleincludes those born between 1950-1972. Restricted sample includes those born between 1957-1962 or 1968-1972. Covariates include: district, year of birth, month of birth, ethnicity (Javanese dummy), birth-yearinteracted with: the number of school-aged children in the district in 1971 (before the start of the program),the enrollment rate of the district in 1971 and the exposure of the district to another INPRES program: awater and sanitation program. Robust standard errors clustered at district of birth. 95% confidence intervalsare included.

77

Figure B.5: Pre-trends raw regression: Second generation outcomes

-.50

.5m

om #

of I

npre

s Sc

hool

s X

yob

1948

1949

1950

1951

1952

1953

1954

1955

1956

1957

1958

1959

1960

1961

1962

mom year of birth

Height-for-age

-.2-.1

0.1

.2.3

mom

# o

f Inp

res

Scho

ols

X yo

b

1948

1949

1950

1951

1952

1953

1954

1955

1956

1957

1958

1959

1960

1961

1962

mom year of birth

if stunted

-.2-.1

0.1

.2m

om #

Inpr

es S

choo

ls X

yob

1948

1949

1950

1951

1952

1953

1954

1955

1956

1957

1958

1959

1960

1961

1962

mom year of birth

Ch sr healthy-.3

-.2-.1

0.1

mom

# In

pres

Sch

ools

X y

ob

1948

1949

1950

1951

1952

1953

1954

1955

1956

1957

1958

1959

1960

1961

1962

mom year of birth

If anemic-.4

-.20

.2.4

mom

# In

pres

Sch

ools

X y

ob

1948

1949

1950

1951

1952

1953

1954

1955

1956

1957

1958

1959

1960

1961

1962

mom year of birth

Poor health Index

-.6-.4

-.20

.2.4

mom

# o

f Inp

res

Scho

ols

X ye

ar o

f birt

h

1948

1949

1950

1951

1952

1953

1954

1955

1956

1957

1958

1959

1960

1961

1962

mom year of birth

Secondary Exam

Mother's exposure -.6

-.4-.2

0.2

.4da

d #

of In

pres

Sch

ools

X y

ob

1948

1949

1950

1951

1952

1953

1954

1955

1956

1957

1958

1959

1960

1961

1962

dad year of birth

Height-for-age

-.10

.1.2

dad

# of

Inpr

es S

choo

ls X

yob

1948

1949

1950

1951

1952

1953

1954

1955

1956

1957

1958

1959

1960

1961

1962

dad year of birth

if stunted

-.1-.0

50

.05

.1da

d #

Inpr

es S

choo

ls X

yob

1948

1949

1950

1951

1952

1953

1954

1955

1956

1957

1958

1959

1960

1961

1962

dad year of birth

Ch sr healthy

-.2-.1

0.1

.2da

d #

Inpr

es S

choo

ls X

yob

1948

1949

1950

1951

1952

1953

1954

1955

1956

1957

1958

1959

1960

1961

1962

dad year of birth

If anemic

-.10

.1.2

dad

# In

pres

Sch

ools

X y

ob

1948

1949

1950

1951

1952

1953

1954

1955

1956

1957

1958

1959

1960

1961

1962

dad year of birth

Poor health Index

-.4-.2

0.2

.4.6

dad

# of

Inpr

es S

choo

ls X

yea

r of b

irth

1948

1949

1950

1951

1952

1953

1954

1955

1956

1957

1958

1959

1960

1961

1962

dad year of birth

Secondary Exam

Father's exposure

Notes: “Ch sr health” corresponds to an indicator of child’s self-reported health status. Coefficients fromdifference-in-differences model that interacts the number of INPRES schools in parent’s birth place andparent’s year of birth. Sample corresponds to children from parent’s born between 1948-1962 (we excludechildren from parents born in 1945-1947 because of very small sample sizes). Bars indicate the 95% confidenceintervals.

78

Figure B.6: Placebo regression: Second generation - each parent separately

HAZ

stunted

SR-Healthy

Anemic

Health Index

Sec. Test Score -Z

-.2 -.1 0 .1 .2SD

Mother Father

Notes: Sample includes the children of individuals born between 1950 and 1962. ”Placebo exposed group”is defined as a dummy equal to one for children born to adults born between 1957 and 1962. Coefficientsreported in standard deviation units. Overall health index corresponds to a summary index from the multiplehealth measures analyzed: height for age z-score, stunting, anemia, self-reported general health. Educationaloutcome is the z-score of the secondary school examination. The health index has mean 0, SD 1. Covariatesinclude the following FE: parent year of birth×1971 enrollment, parent year of birth×1971 number of children,parent year of birth×water sanitation program, child’s gender, birth order, year and month of birth dummies,urban, and ethnicity (Javanese dummy). Because some children may be observed multiple times, robuststandard errors clustered two-way at parent’s district of birth and individual level for health outcomes andclustered at parent’s district of birth for test scores. 95% confidence intervals are included.

79

Table B.1: Replication: Primary completion

Panel A. Expanded sample

(1) (2) (3)

All Male Female

Born between 1963-72 0.028** 0.025* 0.030*

× INPRES (0.014) (0.014) (0.017)

No. of obs. 13,856 6,991 6,865

Dep. var. mean 0.68 0.74 0.62

R-squared 0.252 0.232 0.290

Panel B. Restricted sample

(1) (2) (3)

All Male Female

Born between 1968-72 0.044*** 0.032* 0.052***

× INPRES (0.014) (0.017) (0.018)

No. of obs. 7,650 3,869 3,781

Dep. var. mean 0.73 0.78 0.68

R-squared 0.256 0.271 0.290

Notes: Expanded sample includes those born between 1950-1972. Restricted sample includes those bornbetween 1957-1962 or 1968-1972. Covariates include: district, year of birth, month of birth, ethnicity(Javanese dummy), birth-year interacted with: the number of school-aged children in the district in 1971(before the start of the program), the enrollment rate of the district in 1971 and the exposure of the district toanother INPRES program: a water and sanitation program. Robust standard errors in parentheses clusteredat the district of birth. *** p<0.01, ** p<0.05, * p<0.1

80

Table B.2: First generation health outcomes with multiple hypothesis adjustment

Panel A. Expanded sample

(1) (2)

Original p-value Adjusted p-value

Self-reported healthy 0.000 0.000

Number of days missed 0.013 0.026

Any chronic conditions 0.025 0.031

Number of conditions 0.015 0.026

Mental health 0.202 0.202

Panel B. Restricted sample

(1) (2)

Original p-value Adjusted p-value

Self-reported healthy 0.009 0.047

Number of days missed 0.437 0.437

Any chronic conditions 0.396 0.437

Number of conditions 0.131 0.219

Mental health 0.078 0.196

Notes: Expanded sample includes those born between 1950-1972. Restricted sample includes those bornbetween 1957-1962 or 1968-1972. Covariates include: district, year of birth, month of birth, ethnicity(Javanese dummy), birth-year interacted with: the number of school-aged children in the district in 1971(before the start of the program), the enrollment rate of the district in 1971 and the exposure of the districtto another INPRES program: a water and sanitation program. Robust standard errors clustered at districtof birth. Significance: *** p < 0.01, ** p < 0.05, * p < 0.10.

81

Tab

leB

.3:

Addit

ional

hea

lth

outc

omes

for

the

firs

tge

ner

atio

n

Pan

elA

:E

xpan

ded

sam

ple

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

Hig

hB

Pm

easu

red

Dia

gnos

edhyp

erte

nsi

onB

MI

All

Mal

eF

emal

eA

llM

ale

Fem

ale

All

Mal

eF

emal

eB

orn

bet

.19

63-1

972

0.00

2-0

.011

0.01

2-0

.023

**-0

.013

-0.0

42**

0.16

2*0.

205

0.08

6X

INP

RE

S(0

.009

)(0

.014

)(0

.012

)(0

.011

)(0

.011

)(0

.019

)(0

.095

)(0

.146

)(0

.160

)N

o.of

obs.

1048

050

7254

0810

727

5265

5462

1004

650

1150

35D

ep.

var.

mea

n0.

722

0.71

00.

732

0.25

80.

185

0.32

624

.187

22.9

1525

.460

R-s

quar

ed0.

070.

090.

110.

090.

100.

110.

160.

120.

11G

ender

diff

eren

cep-v

alue

0.20

90.

093

0.58

7

Pan

elB

:R

estr

icte

dsa

mple

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

Hig

hB

Pm

easu

red

Dia

gnos

edhyp

erte

nsi

onB

MI

All

Mal

eF

emal

eA

llM

ale

Fem

ale

All

Mal

eF

emal

eB

orn

bet

.19

68-1

972

-0.0

24**

-0.0

25-0

.015

-0.0

180.

002

-0.0

38**

0.20

1*0.

236*

0.26

9X

INP

RE

S(0

.012

)(0

.018

)(0

.017

)(0

.012

)(0

.012

)(0

.019

)(0

.120

)(0

.143

)(0

.219

)N

o.of

obs.

5837

2807

3030

5939

2900

3039

5639

2781

2858

Dep

.va

r.m

ean

0.69

80.

687

0.70

80.

240

0.15

20.

318

24.4

4023

.172

25.6

31R

-squar

ed0.

080.

120.

130.

110.

120.

140.

160.

130.

14G

ender

diff

eren

cep-v

alue

0.68

80.

031

0.89

7

Not

es:

Exp

and

edsa

mp

lein

clu

des

thos

eb

orn

bet

wee

n1950-1

972.

Res

tric

ted

sam

ple

incl

ud

esth

ose

born

bet

wee

n1957-1

962

or

1968-1

972.

Hig

hb

lood

pre

ssu

red

efin

edas

syst

olic

pre

ssu

regr

eate

rth

an

130

or

dia

stoli

cp

ress

ure

gre

ate

rth

an

80.

Dia

gn

ose

dhyp

erte

nsi

on

base

don

self

rep

ort

edd

iagn

osi

sof

chro

nic

con

dit

ion

s.B

MI

defi

ned

asa

per

son

’sw

eight

inkil

ogra

ms

(kg)

div

ided

by

his

or

her

hei

ght

inm

eter

ssq

uare

d.

Dis

tric

t,ye

ar

of

bir

th,

mon

thof

bir

thfi

xed

effec

tsin

clu

ded

.B

irth

-yea

rin

tera

cted

wit

h:

the

nu

mb

erof

sch

ool-

aged

chil

dre

nin

the

dis

tric

tin

1971

(bef

ore

the

start

of

the

pro

gram

),th

een

roll

men

tra

teof

the

dis

tric

tin

1971

an

dth

eex

posu

reof

the

dis

tric

tto

an

oth

erIN

PR

ES

pro

gra

m:

aw

ate

ran

dsa

nit

ati

on

pro

gra

m.

Rob

ust

stan

dar

der

rors

clu

ster

edat

dis

tric

tof

bir

th.

Sig

nifi

can

ce:

***

p<

0.0

1,

**

p<

0.0

5,

*p<

0.1

0.

82

Table B.4: Gender-specific first generation’s health behavior

Panel A. Expanded sample

(1) (2) (3) (4) (5)

Male Female

Ever Currently No. of daily Alcohol Teen

smoked smoking cigarettes expenditure pregnancy

Born bet. 1963-1972 0.012 0.029** -0.320 779.799* -0.008

X INPRES (0.013) (0.013) (0.380) (408.435) (0.014)

No. of obs. 5296 5296 4041 6006 6003

Dep. var. mean 0.851 0.632 12.439 268.183 0.25

R-squared 0.12 0.11 0.16 0.08 0.14

Panel B. Restricted sample

(1) (2) (3) (4) (5)

Male Female

Ever Currently No. of daily Alcohol Teen

smoked smoking cigarettes expenditure pregnancy

Born bet. 1968-1972 -0.006 0.022 -0.161 446.972 -0.015

X INPRES (0.019) (0.022) (0.560) (500.690) (0.020)

No. of obs. 2930 2930 2242 3407 3310

Dep. var. mean 0.855 0.656 12.550 247.151 0.25

R-squared 0.15 0.14 0.18 0.06 0.16

Notes: Expanded sample includes those born between 1950-1972. Restricted sample includes those bornbetween 1957-1962 or 1968-1972. High blood pressure defined as systolic pressure greater than 130 ordiastolic pressure greater than 80. High BMI defined as BMI greater than 25. District, year of birth, monthof birth fixed effects included. Birth-year interacted with: the number of school-aged children in the districtin 1971 (before the start of the program), the enrollment rate of the district in 1971 and the exposure of thedistrict to another INPRES program: a water and sanitation program. Robust standard errors clustered atdistrict of birth. Significance: *** p < 0.01, ** p < 0.05, * p < 0.10.

83

Tab

leB

.5:

Inte

rgen

erat

ional

outc

omes

:hea

lth

index

-by

gender

Pan

elA

:E

xp

an

ded

Sam

ple

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

Moth

eron

lyF

ath

eron

lyB

oth

pare

nts

All

Son

sD

au

ghte

rsA

llS

on

sD

au

ghte

rsA

llS

on

sD

au

ghte

rs

Mot

her

bet

.19

63-7

2X

INP

RE

S-0

.034***

-0.0

27*

-0.0

47***

-0.0

37**

-0.0

22

-0.0

52**

(0.0

11)

(0.0

15)

(0.0

15)

(0.0

15)

(0.0

21)

(0.0

21)

Fat

her

bet

.19

63-7

2X

INP

RE

S-0

.017

-0.0

07

-0.0

29*

-0.0

03

-0.0

17

0.0

10

(0.0

12)

(0.0

16)

(0.0

16)

(0.0

17)

(0.0

24)

(0.0

22)

No.

ofob

s.17919

9137

8772

17384

8855

8514

19042

9771

9264

R-s

qu

ared

0.1

70.2

00.1

90.1

50.1

80.1

70.1

70.2

00.1

9

Pan

elB

:R

estr

icte

dsa

mp

le

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

Moth

eron

lyF

ath

eron

lyB

oth

pare

nts

All

Son

sD

au

ghte

rsA

llS

on

sD

au

ghte

rsA

llS

on

sD

au

ghte

rs

Mot

her

bet

.19

63-7

2X

INP

RE

S-0

.027*

-0.0

28

-0.0

30

-0.0

26

-0.0

18

-0.0

41

(0.0

16)

(0.0

23)

(0.0

19)

(0.0

19)

(0.0

25)

(0.0

25)

Fat

her

bet

.19

63-7

2X

INP

RE

S-0

.019

-0.0

32

-0.0

19

-0.0

10

-0.0

08

-0.0

09

(0.0

16)

(0.0

21)

(0.0

26)

(0.0

19)

(0.0

27)

(0.0

25)

No.

ofob

s.9911

5068

4829

9007

4627

4357

13707

7073

6617

R-s

qu

ared

0.1

70.2

20.2

10.1

50.1

80.1

80.1

70.2

00.2

0

Note

s:O

ver

all

poor

hea

lth

ind

exco

rres

pon

ds

toa

sum

mary

ind

exfr

om

the

follow

ing

hea

lth

mea

sure

sfo

rth

ese

con

dgen

erati

on

:b

ein

gst

unte

d,

an

emic

an

dse

lf-r

eport

edp

oor

hea

lth

.S

am

ple

corr

esp

on

ds

toch

ild

ren

born

tofi

rst

gen

erati

on

INP

RE

Sin

div

idu

als

.E

xp

an

ded

sam

ple

incl

ud

esch

ild

ren

born

toad

ult

sb

orn

bet

wee

n1950-1

972.

Res

tric

ted

sam

ple

incl

udes

child

ren

born

toad

ult

sb

orn

bet

wee

n1957-1

962

or

1968-1

972.

For

colu

mn

1an

d2,

covari

ate

sin

clu

de

the

foll

ow

ing

FE

:p

are

nt

yea

rof

bir

th×

1971

enro

llm

ent,

pare

nt

yea

rof

bir

th×

1971

nu

mb

erof

child

ren

,p

are

nt

yea

rof

bir

th×

wate

rsa

nit

ati

on

pro

gra

m,

child

’sgen

der

,b

irth

ord

er,

yea

ran

dm

onth

of

bir

thd

um

mie

s,u

rban

,an

det

hn

icit

y(J

avan

ese

du

mm

y).

Rob

ust

stan

dard

erro

rsin

pare

nth

eses

clu

ster

edat

the

pare

nt’

sd

istr

ict

of

bir

th.

For

colu

mn

3(b

oth

pare

nts

),th

esa

mp

leco

rres

pon

ds

toch

ild

ren

eith

erb

orn

moth

ers

or

fath

ers

inth

efi

rst

gen

erati

on

sam

ple

.T

hes

em

od

els

incl

ud

em

oth

er’s

an

dfa

ther

’sex

posu

rean

dth

efu

llse

tof

covari

ate

sfo

rth

em

oth

er,

wh

ile

for

the

fath

erw

ein

clu

de:

pro

vin

ceof

bir

th,

two-y

ear

bin

sfo

ryea

rof

bir

thfi

xed

effec

tsan

din

tera

ctio

ns

bet

wee

nth

efa

ther

’syea

rof

bir

th(i

ntw

o-y

ear

bin

s)an

dth

efa

ther

’sd

istr

ict-

level

covari

ate

s.In

this

esti

mati

on

,st

an

dard

erro

rsare

clu

ster

edtw

o-w

ay

at

the

moth

eran

dfa

ther

dis

tric

tof

bir

th.

Sig

nifi

can

ce:

***

p<

0.0

1,

**

p<

0.0

5,

*p<

0.1

0

84

Tab

leB

.6:

Inte

rgen

erat

ional

outc

omes

:te

stsc

ores

-by

gender

Pan

elA

:E

xp

an

ded

Sam

ple

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

Moth

eron

lyF

ath

eron

lyB

oth

pare

nts

All

Son

sD

au

ghte

rsA

llS

ons

Dau

ghte

rsA

llS

on

sD

au

ghte

rs

Mot

her

bet

.19

63-7

2X

INP

RE

S0.

083***

0.0

91

0.0

63

0.1

10***

0.1

42**

0.0

73

(0.0

31)

(0.0

56)

(0.0

44)

(0.0

39)

(0.0

66)

(0.0

56)

Fat

her

bet

.19

63-7

2X

INP

RE

S0.0

47

0.0

25

0.0

70

0.0

06

-0.0

21

0.0

05

(0.0

33)

(0.0

51)

(0.0

71)

(0.0

43)

(0.0

66)

(0.0

68)

No.

ofob

s.6819

3419

3400

5744

2846

2898

6604

3280

3285

R-s

qu

ared

0.1

10.1

60.1

60.1

10.1

70.1

70.1

40.2

10.1

9

Pan

elB

:R

estr

icte

dsa

mp

le

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

Moth

eron

lyF

ath

eron

lyB

oth

pare

nts

All

Son

sD

au

ghte

rsA

llS

ons

Dau

ghte

rsA

llS

on

sD

au

ghte

rs

Mot

her

bet

.19

63-7

2X

INP

RE

S0.1

05*

0.0

42

0.1

85**

0.1

12**

0.0

32

0.1

78**

(0.0

59)

(0.1

05)

(0.0

73)

(0.0

57)

(0.0

95)

(0.0

82)

Fat

her

bet

.19

63-7

2X

INP

RE

S0.1

03

0.0

20

0.1

39

0.0

41

-0.0

08

-0.0

21

(0.0

65)

(0.1

00)

(0.1

06)

(0.0

56)

(0.0

80)

(0.0

80)

No.

ofob

s.3512

1727

1785

2639

1294

1345

4361

2148

2184

R-s

qu

ared

0.1

40.2

30.2

30.1

70.2

70.3

00.1

70.2

70.2

6

Note

s:O

ver

all

poor

hea

lth

ind

exco

rres

pon

ds

toa

sum

mary

ind

exfr

om

the

follow

ing

hea

lth

mea

sure

sfo

rth

ese

con

dgen

erati

on

:b

ein

gst

unte

d,

an

emic

an

dse

lf-r

eport

edp

oor

hea

lth

.S

am

ple

corr

esp

on

ds

toch

ild

ren

born

tofi

rst

gen

erati

on

INP

RE

Sin

div

idu

als

.E

xp

an

ded

sam

ple

incl

ud

esch

ild

ren

born

toad

ult

sb

orn

bet

wee

n1950-1

972.

Res

tric

ted

sam

ple

incl

udes

child

ren

born

toad

ult

sb

orn

bet

wee

n1957-1

962

or

1968-1

972.

For

colu

mn

1an

d2,

covari

ate

sin

clu

de

the

foll

ow

ing

FE

:p

are

nt

yea

rof

bir

th×

1971

enro

llm

ent,

pare

nt

yea

rof

bir

th×

1971

nu

mb

erof

child

ren

,p

are

nt

yea

rof

bir

th×

wate

rsa

nit

ati

on

pro

gra

m,

child

’sgen

der

,b

irth

ord

er,

yea

ran

dm

onth

of

bir

thd

um

mie

s,u

rban

,an

det

hn

icit

y(J

avan

ese

du

mm

y).

Rob

ust

stan

dard

erro

rsin

pare

nth

eses

clu

ster

edat

the

pare

nt’

sd

istr

ict

of

bir

th.

For

colu

mn

3(b

oth

pare

nts

),th

esa

mp

leco

rres

pon

ds

toch

ild

ren

eith

erb

orn

moth

ers

or

fath

ers

inth

efi

rst

gen

erati

on

sam

ple

.T

hes

em

od

els

incl

ud

em

oth

er’s

an

dfa

ther

’sex

posu

rean

dth

efu

llse

tof

covari

ate

sfo

rth

em

oth

er,

wh

ile

for

the

fath

erw

ein

clu

de:

pro

vin

ceof

bir

th,

two-y

ear

bin

sfo

ryea

rof

bir

thfi

xed

effec

tsan

din

tera

ctio

ns

bet

wee

nth

efa

ther

’syea

rof

bir

th(i

ntw

o-y

ear

bin

s)an

dth

efa

ther

’sd

istr

ict-

level

covari

ate

s.In

this

esti

mati

on

,st

an

dard

erro

rsare

clu

ster

edtw

o-w

ay

at

the

moth

eran

dfa

ther

dis

tric

tof

bir

th.

Sig

nifi

can

ce:

***

p<

0.0

1,

**

p<

0.0

5,

*p<

0.1

0

85

Table B.7: Intergenerational outcomes with multiple hypothesis adjustment

Panel A. Expanded sample

(1) (2) (3) (4)

Maternal exposure Paternal exposure

Original p-value Adjusted p-value Original p-value Adjusted p-value

Height-for-age 0.067 0.084 0.583 0.583

Stunted 0.065 0.084 0.421 0.527

Anemia 0.014 0.036 0.373 0.527

Self-reported healthy 0.348 0.348 0.334 0.527

Secondary test score 0.008 0.036 0.162 0.527

Panel B. Restricted sample

(1) (2) (3) (4)

Maternal exposure Paternal exposure

Original p-value Adjusted p-value Original p-value Adjusted p-value

Height-for-age 0.336 0.444 0.257 0.428

Stunted 0.285 0.444 0.667 0.667

Anemia 0.355 0.444 0.474 0.593

Self-reported healthy 0.459 0.459 0.116 0.291

Secondary test score 0.076 0.378 0.115 0.291

Notes: Expanded sample includes those born between 1950-1972. Restricted sample includes those bornbetween 1957-1962 or 1968-1972. Covariates include: district, year of birth, month of birth, ethnicity(Javanese dummy), birth-year interacted with: the number of school-aged children in the district in 1971(before the start of the program), the enrollment rate of the district in 1971 and the exposure of the districtto another INPRES program: a water and sanitation program. Robust standard errors clustered at districtof birth. Significance: *** p < 0.01, ** p < 0.05, * p < 0.10.

86

Table B.8: Intergenerational outcomes: average poor health index

Panel A. Expanded sample

(1) (2) (3)

Mother only Father only Both parents

Mother bet. 1963-72 X INPRES -0.034*** -0.026

(0.013) (0.017)

Father bet. 1963-72 X INPRES -0.018 -0.008

(0.013) (0.018)

No. of obs. 10712 10805 11703

R-squared 0.19 0.17 0.19

Panel B. Restricted sample

(1) (2) (3)

Mother only Father only Both parents

Mother bet. 1963-72 X INPRES -0.021 -0.016

(0.020) (0.021)

Father bet. 1963-72 X INPRES -0.020 -0.008

(0.019) (0.020)

No. of obs. 6084 5686 8531

R-squared 0.20 0.17 0.20

Notes: Sample includes the first generation’s children who are between ages 8 and 18. Sample correspondsto children born to first generation INPRES individuals. Expanded sample includes children born to adultsborn between 1950-1972. Restricted sample includes children born to adults born between 1957-1962 or1968-1972. Covariates include the following FE: parent year of birth and district of birth fixed effects, parentyear of birth×1971 enrollment, parent year of birth×1971 number of children, parent year of birth×watersanitation program, child’s gender, birth order, year and month of birth dummies, ethnicity (Javanesedummy), examination year dummies. Average regressions weighted by the number of observations per child.Robust standard errors in parentheses clustered at the parent’s district of birth. *** p<0.01, ** p<0.05, *p<0.1

87

Table B.9: Intergenerational outcomes: Complete secondary school

Panel A. Expanded sample

(1) (2) (3)

Mother only Father only Both parents

Mother bet. 1963-72 X INPRES -0.007 -0.012

(0.010) (0.014)

Father bet. 1963-72 X INPRES 0.015 0.021

(0.013) (0.015)

No. of obs. 13446 10764 12466

Dep. var. mean 0.76 0.78 0.76

R-squared 0.23 0.24 0.25

Panel B. Restricted sample

(1) (2) (3)

Mother only Father only Both parents

Mother bet. 1963-72 X INPRES 0.005 -0.011

(0.014) (0.017)

Father bet. 1963-72 X INPRES 0.018 0.024

(0.021) (0.017)

No. of obs. 6724 4725 8040

Dep. var. mean 0.77 0.78 0.77

R-squared 0.26 0.27 0.27

Notes: Sample corresponds to children born to first generation INPRES individuals. Expanded sampleincludes children born to adults born between 1950-1972. Restricted sample includes children born to adultsborn between 1957-1962 or 1968-1972. Covariates include the following FE: parent year of birth and districtof birth fixed effects, parent year of birth×1971 enrollment, parent year of birth×1971 number of children,parent year of birth×water sanitation program, child’s gender, birth order, year and month of birth dummies,ethnicity (Javanese dummy). Robust standard errors in parentheses clustered at the parent’s district of birth.Significance: *** p < 0.01, ** p < 0.05, * p < 0.10.

88

Table B.10: Robustness: Alternative exposure based on schools built per year

(1) (2) (3) (4) (5) (6) (7)

First generation Second generation

Health index Health index Test score

Maternal Paternal Maternal Paternal

All Male Female exposure exposure exposure exposure

Born bet. 1963-1972 -0.032** -0.027 -0.035* -0.012 -0.014 0.116*** 0.104**

× INPRES (0.013) (0.019) (0.021) (0.013) (0.014) (0.038) (0.045)

No. of obs. 9836 4785 5051 17851 17306 6802 5718

R-squared 0.10 0.12 0.11 0.17 0.15 0.11 0.11

Notes: The alternative exposure variable is based on the number of schools built per year for each cohort.

Expanded sample includes children born to adults born between 1950-1972. Restricted sample includes

children born to adults born between 1957-1962 or 1968-1972. See Table 1, Table 3, and Table 5 for

covariates. Robust standard errors clustered at district of birth for cols. 1-3. Robust standard errors in

parentheses clustered at the parent’s district of birth for cols. 4-5. Robust standard errors in parentheses

two-way clustered at the parent’s district of birth and individual level for cols. 6-7. Significance: *** p <

0.01, ** p < 0.05, * p < 0.10.

89

Table B.11: Robustness: Alternative cohorts

Panel A. Expanded sample

(1) (2) (3) (4) (5) (6) (7)

First generation Second generation

Health index Health index Test score

All Male Female Maternal Paternal Maternal Paternal

Born bet. 1963-1975 -0.041*** -0.021 -0.058*** -0.024** -0.013 0.102*** 0.041

× INPRES (0.013) (0.021) (0.019) (0.012) (0.011) (0.029) (0.034)

No. of obs. 9891 4817 5074 20341 19145 7338 5981

R-squared 0.10 0.12 0.11 0.16 0.14 0.11 0.11

Panel B. Restricted sample

(1) (2) (3) (4) (5) (6) (7)

First generation Second generation

Health index Health index Test score

All Male Female Maternal Paternal Maternal Paternal

Born bet. 1968-1975 -0.038** -0.010 -0.061** -0.014 -0.012 0.133*** 0.086

× INPRES (0.016) (0.032) (0.026) (0.015) (0.014) (0.050) (0.061)

No. of obs. 5537 2668 2869 12334 10766 4031 2876

R-squared 0.12 0.15 0.14 0.16 0.13 0.13 0.15

Notes: Expanded sample includes those born between 1950-1975. Restricted sample includes those born

between 1957-1962 or 1968-1975. See Table 1 , Table 3, and Table 5 for covariates. Robust standard errors

clustered at district of birth for cols. 1-3. Robust standard errors in parentheses clustered at the parent’s

district of birth for cols. 4-5. Robust standard errors in parentheses two-way clustered at the parent’s district

of birth and individual level for cols. 6-7. Significance: *** p < 0.01, ** p < 0.05, * p < 0.10.

90

Table B.12: Potential mechanism: Household resources

Panel A. Expanded sample

(1) (2) (3) (4) (5) (6)

Food Non-food

All Male Female All Male Female

Born bet. 1963-1972 0.040*** 0.038* 0.041 0.053** 0.050 0.064

× INPRES (0.015) (0.022) (0.027) (0.026) (0.034) (0.043)

No. of obs. 11941 6007 5934 11925 5998 5927

Dep. var. mean 12.430 12.481 12.378 11.522 11.577 11.467

R-squared 0.14 0.15 0.17 0.16 0.16 0.20

Panel B. Restricted sample

(1) (2) (3) (4) (5) (6)

Food Non-food

All Male Female All Male Female

Born bet. 1968-1972 0.040* 0.052 0.031 0.069** 0.062 0.105**

× INPRES (0.023) (0.032) (0.034) (0.034) (0.048) (0.053)

No. of obs. 6752 3408 3344 6742 3403 3339

Dep. var. mean 12.458 12.499 12.417 11.584 11.622 11.545

R-squared 0.14 0.15 0.19 0.16 0.17 0.22

Notes: Log per capita expenditure in 2012-14 based on weekly or monthly per capita food and non-foodexpenditure in 2012 Rupiah). Expanded sample includes those born between 1950-1972. Restricted sampleincludes those born between 1957-1962 or 1968-1972. See Table 1 for covariates. Robust standard errorsclustered at district of birth. Significance: *** p < 0.01, ** p < 0.05, * p < 0.10.

91

Table B.13: Potential mechanism: Housing quality (Expanded sample)

(1) (2) (3) (4) (5) (6)

Drinking Bad Bad Bad Bad Bad

water Toilet Occupancy Floor Roof Wall

Panel A. All

Born between 1963-1972 0.010 -0.006 -0.015* -0.006 0.003 -0.012

× INPRES (0.009) (0.007) (0.008) (0.006) (0.003) (0.008)

N 11835 11849 11934 12132 12121 12022

Dep. var. mean 0.437 0.086 0.092 0.163 0.033 0.234

R-squared 0.21 0.13 0.11 0.39 0.27 0.33

Panel B. Men

Born between 1963-1972 0.010 0.001 -0.026** -0.010 0.000 -0.009

× INPRES (0.013) (0.009) (0.012) (0.008) (0.003) (0.010)

N 5887 5891 5944 6042 6036 5986

Dep. var. mean 0.435 0.087 0.094 0.164 0.031 0.235

R-squared 0.24 0.15 0.14 0.43 0.27 0.36

Panel C. Women

Born between 1963-1972 0.013 -0.011 -0.002 -0.002 0.004 -0.019*

× INPRES (0.012) (0.009) (0.008) (0.009) (0.005) (0.011)

N 5948 5958 5990 6090 6085 6036

Dep. var. mean 0.439 0.086 0.089 0.162 0.035 0.233

R-squared 0.22 0.15 0.12 0.38 0.31 0.34

Notes: Poor toilet is captured by not having access to a toilet (including shared or public toilet). Poor floorincludes board or lumber, bamboo, or dirt floor. Poor roof includes leaves or wood. Poor wall includeslumber or board and bamboo or mat. High occupancy per room is defined as more than two persons perroom in the house (based on household size). Expanded sample includes those born between 1950-1972. SeeTable 1 for covariates. Robust standard errors clustered at district of birth. Significance: *** p < 0.01, **p < 0.05, * p < 0.10.

92

Table B.14: Potential mechanism: Housing quality (Restricted sample)

(1) (2) (3) (4) (5) (6)

Drinking Bad Bad Bad Bad Bad

water Toilet Occupancy Floor Roof Wall

Panel A. All

Born between 1968-1972 0.016 -0.003 -0.012 -0.006 0.002 -0.020**

× INPRES (0.013) (0.008) (0.010) (0.008) (0.004) (0.009)

N 6696 6698 6760 6867 6861 6795

Dep. var. mean 0.449 0.085 0.092 0.161 0.035 0.232

R-squared 0.22 0.14 0.11 0.39 0.27 0.34

Panel B. Men

Born between 1968-1972 0.006 -0.002 -0.037*** -0.015 0.005 -0.026**

× INPRES (0.016) (0.010) (0.013) (0.010) (0.005) (0.011)

N 3319 3319 3363 3408 3405 3371

Dep. var. mean 0.448 0.087 0.096 0.160 0.032 0.231

R-squared 0.25 0.17 0.15 0.43 0.27 0.37

Panel C. Women

Born between 1968-1972 0.043** -0.006 0.013 0.000 -0.001 -0.020

× INPRES (0.020) (0.012) (0.014) (0.012) (0.007) (0.014)

N 3377 3379 3397 3459 3456 3424

Dep. var. mean 0.450 0.082 0.088 0.162 0.037 0.234

R-squared 0.24 0.17 0.14 0.40 0.33 0.35

Notes: See Table B.13 for notes. Restricted sample includes those born between 1957-1962 or 1968-1972.Robust standard errors clustered at district of birth. Significance: *** p < 0.01, ** p < 0.05, * p < 0.10.

93

Table B.15: Robustness: Non-movers only

Panel A. Expanded sample

(1) (2) (3) (4) (5) (6) (7)

First generation Second generation

Health index Health index Test score

All Male Female Maternal Paternal Maternal Paternal

Born bet. 1963-1972 -0.039* 0.008 -0.074*** -0.037** -0.030** 0.100*** 0.032

× INPRES (0.023) (0.033) (0.028) (0.014) (0.015) (0.038) (0.036)

No. of obs. 4779 2271 2508 12738 11568 4731 3790

R-squared 0.12 0.15 0.14 0.18 0.16 0.12 0.14

Panel B. Restricted sample

(1) (2) (3) (4) (5) (6) (7)

First generation Second generation

Health index Health index Test score

All Male Female Maternal Paternal Maternal Paternal

Born bet. 1968-1972 -0.026 0.019 -0.052 -0.044** -0.032* 0.142* 0.114

× INPRES (0.025) (0.047) (0.032) (0.017) (0.016) (0.076) (0.071)

No. of obs. 2662 1275 1387 6987 5941 2476 1778

R-squared 0.13 0.16 0.18 0.18 0.16 0.17 0.20

Notes: Expanded sample includes those born between 1950-1975. Restricted sample includes those bornbetween 1957-1962 or 1968-1975. See Table 1, Table 3, and Table 5 for covariates. Robust standard errorsclustered at district of birth for cols. 1-3. Robust standard errors in parentheses clustered at the parent’sdistrict of birth for cols. 4-5. Robust standard errors in parentheses two-way clustered at the parent’s districtof birth and individual level for cols. 6-7. Significance: *** p < 0.01, ** p < 0.05, * p < 0.10.

94

Tab

leB

.16:

Pot

enti

alm

echan

ism

:N

eigh

bor

hood

qual

ity

(Expan

ded

sam

ple

)

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

(10)

Ele

men

tary

Mid

dle

Hig

hP

iped

Pri

vate

Clinic

Mid

wiv

esR

ice

Nat

ional

Dis

tric

tsc

hool

school

school

wat

erpro

vid

ers

subsi

dy

hea

lth

insu

rance

hea

lth

insu

rance

Pan

elA

.A

llB

orn

bet

wee

n19

63-1

972

0.00

10.

028

0.04

40.

010

0.00

40.

010

0.02

8-0

.000

0.00

10.

004

×IN

PR

ES

(0.0

90)

(0.0

59)

(0.0

60)

(0.0

11)

(0.0

06)

(0.0

25)

(0.0

25)

(0.0

05)

(0.0

06)

(0.0

06)

N93

2993

2993

2991

4093

2993

2992

6790

6385

3883

94D

ep.

var.

mea

n5.

466

4.45

14.

735

0.34

20.

863

2.11

01.

205

0.32

60.

328

0.26

1R

-squar

ed0.

550.

490.

480.

430.

510.

420.

440.

470.

420.

42

Pan

elB

.M

enB

orn

bet

wee

n19

63-1

972

0.11

40.

134

0.12

30.

022

-0.0

030.

037

0.04

70.

005

0.00

80.

011

×IN

PR

ES

(0.1

29)

(0.1

05)

(0.1

04)

(0.0

19)

(0.0

08)

(0.0

31)

(0.0

32)

(0.0

09)

(0.0

11)

(0.0

11)

N45

2645

2645

2644

3045

2645

2644

9543

9941

5640

99D

ep.

var.

mea

n5.

405

4.39

74.

679

0.33

10.

853

2.09

91.

213

0.33

10.

331

0.26

0R

-squar

ed0.

570.

500.

500.

430.

530.

440.

470.

480.

430.

42

Pan

elC

.W

omen

Bor

nb

etw

een

1963

-197

2-0

.035

-0.0

170.

018

-0.0

040.

009

0.00

10.

004

-0.0

08-0

.003

-0.0

03×

INP

RE

S(0

.125

)(0

.068

)(0

.083

)(0

.013

)(0

.009

)(0

.036

)(0

.029

)(0

.007

)(0

.006

)(0

.008

)N

4776

4776

4776

4683

4776

4776

4744

4637

4353

4269

Dep

.va

r.m

ean

5.51

94.

507

4.79

50.

352

0.87

42.

120

1.20

00.

322

0.32

60.

263

R-s

quar

ed0.

560.

520.

500.

470.

520.

440.

460.

490.

440.

46

Not

es:

Poor

toil

etis

cap

ture

dby

not

hav

ing

acce

ssto

ato

ilet

(incl

ud

ing

share

dor

pu

bli

cto

ilet

).P

oor

floor

incl

ud

esb

oard

or

lum

ber

,b

am

boo,

or

dir

tfl

oor

.P

oor

roof

incl

ud

esle

aves

orw

ood

.P

oor

wall

incl

ud

eslu

mb

eror

board

an

db

am

boo

or

mat.

Hig

hocc

up

an

cyp

erro

om

isd

efin

edas

more

than

two

per

son

sp

erro

omin

the

hou

se(b

ased

onh

ou

seh

old

size

).E

xp

an

ded

sam

ple

incl

ud

esth

ose

born

bet

wee

n1950-1

972.

See

Tab

le1

for

cova

riate

s.R

obu

stst

and

ard

erro

rscl

ust

ered

atd

istr

ict

ofb

irth

.S

ign

ifica

nce

:***

p<

0.0

1,

**

p<

0.0

5,

*p<

0.1

0.

95

Tab

leB

.17:

Pot

enti

alm

echan

ism

:N

eigh

bor

hood

qual

ity

(Res

tric

ted

sam

ple

)

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

(10)

Ele

men

tary

Mid

dle

Hig

hP

iped

Pri

vate

Clinic

Mid

wiv

esR

ice

Nat

ional

Dis

tric

tsc

hool

school

school

wat

erpro

vid

ers

subsi

dy

hea

lth

insu

rance

hea

lth

insu

rance

Pan

elA

.A

llB

orn

bet

wee

n19

63-1

972

-0.0

65-0

.012

0.03

40.

017

0.00

80.

061*

0.04

80.

003

-0.0

07-0

.007

×IN

PR

ES

(0.1

10)

(0.0

72)

(0.0

93)

(0.0

12)

(0.0

10)

(0.0

35)

(0.0

31)

(0.0

08)

(0.0

06)

(0.0

09)

N51

1851

1851

1850

1551

1851

1850

7749

5946

7245

99D

ep.

var.

mea

n5.

514

4.42

14.

670

0.35

40.

861

2.09

21.

216

0.32

30.

329

0.26

6R

-squar

ed0.

570.

520.

490.

460.

500.

440.

470.

460.

420.

42

Pan

elB

.M

enB

orn

bet

wee

n19

63-1

972

0.11

7-0

.024

0.11

00.

018

-0.0

020.

060

0.03

10.

014

0.00

20.

009

×IN

PR

ES

(0.1

62)

(0.1

04)

(0.1

06)

(0.0

18)

(0.0

09)

(0.0

40)

(0.0

43)

(0.0

11)

(0.0

11)

(0.0

14)

N24

6424

6424

6424

1124

6424

6424

4223

8722

5122

21D

ep.

var.

mea

n5.

419

4.33

94.

635

0.33

80.

852

2.06

41.

226

0.32

30.

329

0.27

2R

-squar

ed0.

600.

510.

520.

470.

520.

440.

490.

490.

450.

45

Pan

elC

.W

omen

Bor

nb

etw

een

1963

-197

2-0

.117

0.03

30.

047

0.02

20.

017

0.06

70.

067*

-0.0

13-0

.017

**-0

.019

×IN

PR

ES

(0.1

60)

(0.1

15)

(0.1

40)

(0.0

15)

(0.0

13)

(0.0

56)

(0.0

40)

(0.0

11)

(0.0

08)

(0.0

16)

N26

1726

1726

1725

6826

1726

1725

9825

3423

8323

41D

ep.

var.

mea

n5.

596

4.49

34.

702

0.36

50.

869

2.11

71.

211

0.32

40.

328

0.25

9R

-squar

ed0.

590.

570.

500.

490.

520.

480.

500.

480.

430.

46

Not

es:

Poor

toil

etis

cap

ture

dby

not

hav

ing

acce

ssto

ato

ilet

(incl

ud

ing

share

dor

pu

bli

cto

ilet

).P

oor

floor

incl

ud

esb

oard

or

lum

ber

,b

am

boo,

or

dir

tfl

oor

.P

oor

roof

incl

ud

esle

aves

orw

ood

.P

oor

wall

incl

ud

eslu

mb

eror

board

an

db

am

boo

or

mat.

Hig

hocc

up

an

cyp

erro

om

isd

efin

edas

more

than

two

per

son

sp

erro

omin

the

hou

se(b

ased

onh

ou

seh

old

size

).E

xp

an

ded

sam

ple

incl

ud

esth

ose

born

bet

wee

n1950-1

972.

See

Tab

le1

for

cova

riate

s.R

obu

stst

and

ard

erro

rscl

ust

ered

atd

istr

ict

ofb

irth

.S

ign

ifica

nce

:***

p<

0.0

1,

**

p<

0.0

5,

*p<

0.1

0.

96

Tab

leB

.18:

Pot

enti

alm

echan

ism

san

dth

eir

contr

ibuti

ons

(1)

(2)

(3)

(4)

(5)

(6)

(7)

Ass

oci

atio

nE

xp

an

ded

sam

ple

Res

tric

ted

sam

ple

bet

wee

nch

an

nel

INP

RE

Seff

ect

INP

RE

Seff

ect

Contr

ibu

tion

INP

RE

Seff

ect

INP

RE

Seff

ect

Contr

ibu

tion

and

outc

ome

on

chan

nel

on

ou

tcom

e(%

)on

chan

nel

on

ou

tcom

e(%

)

Pan

elA

:F

irst

gen

erati

on

Outcome:

Poorhealthindex

Log

per

cap

exp

endit

ure

s0.

016

0.0

46

0.0

41.8

40%

0.0

54

0.0

33

2.6

18%

Nei

ghb

orh

ood

qu

alit

y:

Hea

lth

reso

urc

es0.

018

0.0

24

0.0

41.0

80%

0.0

55

0.0

33

3.0

00%

Ass

orta

tive

mat

ing

for

wom

en0.

025

0.0

06

0.0

40.3

82%

0.0

32

0.0

33

0.0

81%

Pan

elB

:S

econ

dgen

erati

on

Outcome:

Poorhealthindex

Nu

mb

erof

chil

dre

n0.

019

-0.0

48

-0.0

34

2.7

23%

-0.0

52

-0.0

27

3.7

14%

Log

per

cap

exp

endit

ure

s-0

.045

0.0

46

-0.0

34

6.1

31%

0.0

54

-0.0

27

9.0

63%

Nei

ghb

orh

ood

qu

alit

y:

Hea

lth

reso

urc

es-0

.036

0.0

24

-0.0

34

2.5

68%

0.0

55

-0.0

27

7.4

10%

Par

ents

’ed

uca

tion

:M

oth

er-0

.071

0.0

3-0

.034

6.2

48%

0.0

52

-0.0

27

13.6

37%

Fat

her

-0.0

490.0

25

-0.0

34

3.6

07%

0.0

32

-0.0

27

5.8

13%

Outcome:

Heightforage

z-score

Nu

mb

erof

chil

dre

n-0

.072

-0.0

48

0.0

56

6.1

72%

-0.0

52

0.0

44

8.5

10%

Log

per

cap

exp

endit

ure

s0.

179

0.0

46

0.0

56

14.7

27%

0.0

54

0.0

44

22.0

03%

Nei

ghb

orh

ood

qu

alit

y:

Hea

lth

reso

urc

es0.

079

0.0

24

0.0

56

3.3

70%

0.0

55

0.0

44

9.8

28%

Par

ents

’ed

uca

tion

:M

oth

er0.

163

0.0

30.0

56

8.7

44%

0.0

52

0.0

44

19.2

89%

Fat

her

0.08

10.0

25

0.0

56

3.6

08%

0.0

32

0.0

44

5.8

77%

Outcome:

Testscores

Nu

mb

erof

chil

dre

n-0

.040

-0.0

48

0.0

83

2.2

85%

-0.0

52

0.1

05

1.9

56%

Log

per

cap

exp

endit

ure

s0.

117

0.0

46

0.0

83

6.4

84%

0.0

54

0.1

05

6.0

17%

Par

ents

’ed

uca

tion

:M

oth

er0.

177

0.0

30.0

83

6.4

09%

0.0

52

0.1

05

8.7

82%

Fat

her

0.11

80.0

25

0.0

83

3.5

49%

0.0

32

0.1

05

3.5

91%

Note

s:E

xp

an

ded

sam

ple

incl

ud

esth

ose

born

bet

wee

n1950-1

972.

See

Tab

le1

for

covari

ate

s.R

ob

ust

stan

dard

erro

rscl

ust

ered

at

dis

tric

tof

bir

th.

Sig

nifi

can

ce:

***

p<

0.0

1,

**

p<

0.0

5,

*p<

0.1

0.

97

C Cost-Benefit Analysis

Cost

We follow Duflo (2001) in our cost estimation. The cost of the school construction and the

number of schools built each year came from Duflo (2001). We assume the schools were

built between 1973 and 1977 and were operational for 20 years, so the school’s last year of

operation is 1997. Each school was designed with 3 classrooms and 3 teachers. Teacher’s

salary was USD 360 (in 1990 USD) in 1974 and USD 2467 (in 1990 USD) in 1995, and we

assume linear growth between 1974 and 1995. We assume each teacher would require

training, and that would cost a third of the salary. The maintenance cost is assumed to be

25% of the wage bill.

Cohort size

We estimate the number of first generation individuals exposed to the program, starting

from those born in 1963. Assuming children start school at 6, the last cohort to benefit is

born in 1989. We use the 1971, 1980, and 1990 Census to obtain the population of the

cohort born between 1963 and 1989. We then estimate the number of students enrolled

based on the enrollment rates between 1970 and 1995 from the World Development

Indicators.

The program sought to attain a teacher student ratio of 40 students per teacher. Each

school would typically hold a morning and afternoon session, with 3 classes in each session,

so each school served 240 students.58 To obtain the INPRES coverage for each year, we use

the number of INPRES schools in each year divided by the number of primary school-aged

children (6-12 year olds) enrolled in primary school. The INPRES exposure for each cohort

is given by the average INPRES coverage for the cohort’s primary schooling (6 years).

First generation benefits

We follow the literature and assume that returns to primary education is 20%

(Psacharopoulos and Patrinos*, 2004). To calculate the base earnings on which to apply

the return to primary schooling, we assume the Indonesian population will be in the labor

force between the ages of 16 and 55, which is the official age of retirement up to the early

2000s. We calculate the mean earnings of individuals aged 16 to 55 in IFLS-1. We then use

the CPI to deflate earnings to 1990 USD. The estimated lifetime earnings is about USD

57000 (in 1990 USD).58Conversations with Bappenas and former Bappenas officials.

98

Figure B.7: Cost Benefit Analysis

Notes: We assume the benefits are derived solely from the earnings gain of the first generation individuals.

Cohorts born in 1963 to 1989 benefit from the program. Benefits accrue from 1979, the first year that the

first cohort entered the labor market, and end in 2052, when the last cohort served by the program would

retire.

We include the gains from health based on the relationship between poor health and

mortality at older ages. We follow the literature and assume that self-reported poor health

is associated with a 2.73 odds ratio among those 50 and above in Indonesia (Frankenberg

and Jones, 2004). We then combine this with mean earnings between the ages of 50 and 55

(in 1990 USD) and estimated survival probability for those ages from Statistics Indonesia.59

Second generation benefits

To obtain the cohort size in the second generation, we assume each first generation

individual has 1.2 children at age 22.60 We assume second generation individuals have 20%

higher lifetime earnings compared to the first generation individuals and the second

generation would be in the labor market between the ages of 16 and 55. The effect of

INPRES on the second generation’s height is 0.056 standard deviations. With about 6

centimeters standard deviation in height, this would correspond to about 0.366 centimeters

height increase. With an 8% gain in earnings resulting from the height premium (Sohn,

2015), the program effect would then translate to a 0.26% gain in lifetime earnings for the

second generation. For the second generation gain in education, we use literacy as a proxy

59Pengembangan Model Life Table Indonesia (2011). Last accessed July 15, 2019.60Indonesia’s total fertility rate in 2000 is 2.4 per woman, so we assume a fertility rate of 1.2 for the first

generation individuals.

99

Table B.19: Internal Rate of Return Estimates

Internal Rate of Return

First generation Earnings returns to Primary Completion 7.91%+ earnings gains from better health and lower mortality 8.75%and lower mortality

Second generation 1st gen. gains + returns to height 14.53%1st gen. gains + returns to test scores 21.48%1st gen. gains + returns to height and test scores 24.76%if independent

Notes: First generation earnings based on returns to primary completion. First generation health gainsbased on earnings gains between 50 and 55 from program reduction in poor health that is associated withmortality improvement. Test score gains based on earnings gain from improved secondary test score. Healthgains based on the height premium. Test score and health are assumed to be independent.

for gains in test score. Following the literature, we assume a one standard deviation

increase in literacy would increase earnings by 8.5% (Perez-Alvarez, 2017). The program

effect would then translate to a 0.68% gain in lifetime earnings for the second generation.

Scenarios

We present calculations based on several scenarios (Table B.19). First, we assume the gains

for the first generation came from earnings only. Next, we assume the gains for the first

generation came from earnings and mortality gains. We then add the gains from the

second generation’s test score alone, the second generation’s health alone, and finally, we

assume the gains from health and test scores are independent and combine the gains.

100